AI commentators are now calling for ‘intent engineering’, making organizational goals, trade-off hierarchies and decision boundaries explicit enough that AI agents can actually act on them. The problem is that nobody has specified what that looks like in practice.

DoView outcomes modeling, a methodology developed and applied in thousands of human organizational contexts over more than two decades, may already be the answer. DoViews distil strategic intent into a single structured visual artifact that serves as a shared source of truth for execution. What they have done for human organizations, getting intent out of people’s heads and into a form that can guide and regulate action, is potentially directly transferable to AI intent engineering. This article sets out how, including a preliminary test using the tau2-bench benchmark, and a concrete proposal for a two-version DoView: one human-readable, one machine-actionable.

Introduction

Well-known AI commentator Nate Jones, in a recent widely circulated video on AI in the enterprise, focuses on the term ‘intent engineering’ and describes it as the most important challenge in deploying AI agents at organisational scale. He is right. He is also, without realising it, describing DoView outcomes modelling.

Nate states his argument most clearly here:

‘This means that the most important AI investment in 2026 isn’t really a model subscription. It’s not another Copilot license. It’s organizational intent architecture, making your company’s goals, values, decision frameworks, and trade off hierarchies discoverable, structured, and agent actionable. It’s building the alignment infrastructure that lets agents make decisions that aren’t just technically correct, but that are strategically coherent.’

This is exactly what DoView outcomes diagrams are designed for. Used in thousands of cases, they are built according to a standard for capturing and visualising organisational intent, and the methodology can be directly applied to intent engineering for AI systems.

Just for a starter, to quickly emphasise the main point of this article, the front image of Nate’s video on this topic even resembles a DoView outcomes diagram. This is no accident. DoViews are a tried-and-tested tool that provides the basis for a practical ‘intent methodology’ which Nate is seeking.

Nate’s recent video on intent engineering

A DoView outcomes diagram for an AI system

First, what is a DoView outcomes diagram?

Nate thinks that we need the following:

‘Making organizational intent explicit and structured is extremely difficult. Most organizations have never had to do this. Their goals live in slide decks in OKR [Objectives and Key Results] documents that get half read and referenced at personal reviews once a year, in leadership principles that get cited in performance reviews, but, really, they don’t get operationalized.’

The purpose of DoView outcomes diagrams within DoView Planning is to do exactly what Nate is calling for here. The only point I want to differ on with him is where he says that it is extremely difficult to make organizational intent explicit. In fact, using DoView outcomes diagrams, it is now extremely easy to do this. Well, at least to develop a draft with the DoView drawing prompt. Then you just get the organization to review the draft to make sure that it captures their organizational intent. I have done this successfully with hundreds of organizations. They have then gone on to use the DoView outcomes diagrams created as the basis for their planning, implementation and reporting.

So, DoView outcomes diagrams can easily capture organizational intent. They are created following a standard (the DoView drawing rules) to capture the essential strategic intent of any organisation in an accessible and human-readable format. A DoView distils strategic intent from an organisation’s strategic documentation plus the mental models of strategy held by leadership, to serve as a single source of strategic truth for the organization. The DoView is then utilised at all levels within the organisation to ensure alignment.

You can view DoViews for companies such as Anthropic, Apple, Google, JP Morgan Chase, KPMG US, Microsoft, Nike, Open AI, and Tesla. Or see DoViews for smaller organizations. Anyone can also, for free, build a PowerPoint version of a DoView for any organisation or initiative involving humans or AI agents, using the DoView drawing prompt.

Before we get into the details of what Nate is arguing for, his call for intent engineering aligns with a Microsoft report from last year. The report says that with the introduction of AI, we need to shift from thinking in terms of organizational charts to a new concept of ‘outcomes-driven work charts’. The report did not detail what these look like, but consistent with the Microsoft call, this describes exactly what DoView outcomes diagrams are.

What Nate is saying regarding intent engineering

Nate’s argument is that the AI race is no longer an intelligence race. Rather, it is an ‘intent race’. The frontier models, Opus, Gemini, GPT-5, are all very capable. The gap between them matters far less than the gap between organizations that give those models clear, structured, goal-aligned intent and organizations that do not.

He puts it like this:

‘The company with a mediocre model and extraordinary organizational intent infrastructure will outperform the company with a frontier model and fragmented, inaccessible, unaligned organizational knowledge every single time.’

Nate outlines three layers of what he considers the AI maturity stack. At the top, the layer he says almost certainly does not exist in current organizations is intent engineering.

Here is his definition of the problem he wants to address:

‘An agent does not know your company’s OKRs [Objectives and Key Results] unless you put them in the context window. It doesn’t know which trade-offs your leadership team would prefer unless you encode those preferences in a way it can act on. It doesn’t know the difference between a decision that should be escalated and one it should make autonomously unless you define the boundary.’

This is exactly what is spelt out in the human organizational context in the form of a DoView outcomes diagram. All of the strategic planning, measurement and risk documentation for any organization is distilled down into a single DoView outcomes model, which is then the vehicle used to communicate the distilled strategy throughout the organization as it executes on its intent.

The Klarna Lesson Nate Keeps Coming Back To

Nate repeatedly returns to the Klarna story as the cautionary tale of what happens when you deploy AI agents without intent engineering. Klarna replaced hundreds of customer service agents with AI, then had to walk back that decision and rehire. He explains what went wrong:

‘The agent optimized for resolution speed because that was the objective it could measure. Nobody had encoded the objectives that mattered most, relationship quality, brand trust, customer lifetime value, the contextual judgment about when to be efficient and when to be generous. Those objectives lived in the heads of the human agents who had to walk out the door.’

This is exactly the point that is made repeatedly in DoView Planning. A key function of visualizing organizational intent as DoView outcomes models is to provide an external artefact that can be used as a shared thinking tool for all levels of execution. Strategic intent in organizations, and now when trying to get AI agents to do what we want them to do, cannot just be left residing in people’s heads. While it stays in people’s heads, there is no way to know whether they have a fully articulated model of the strategy that they are trying to execute.

The direct analogy here is that you would not leave what you want architects and builders to do just inside their heads when constructing a high-rise. Instead, intent is captured in a visual artefact in the external world, a blueprint that is then used to guide and regulate all discussions about the construction. Similarly, generals planning a battle do not keep the map of the territory just in their minds. It is likewise represented as a real-world visual artefact that they collectively examine to make strategic decisions.

Exactly the same reasoning applies to taking any other type of action in the world. You need to get strategic intent out of people’s heads and into a concrete, real-world artifact that they can use for joint decision making. This is the exact purpose of a DoView outcomes model.

What DoView models have done for human organizations can now be done for AI systems

What DoView outcomes models have done for human organizations is now directly transferable to answering Nate’s call for the specification of intent when building and managing AI agent systems. The obvious answer in regard to intent engineering for AI systems is to do exactly what has been done thousands of times with DoView outcomes diagrams for human organizations. Extract the mental model, represent it visually in a standard format, check that all the parties involved have the same mental model and then use it to execute the action that you want undertaken. How DoView diagrams are used in human organization planning is set out here.

In the Klarna case Nate talks about, the agent was not misaligned because the technology failed. It was misaligned because the implicit DoView it was working from was wrong. The box that said ‘resolve calls faster’ was in it. The boxes that said ‘preserve customer lifetime value’ and ‘maintain brand trust’ were not. Having a DoView of Klarna’s customer service agent at the time of deployment would have made that gap immediately visible. Anyone reviewing the DoView the agent was working off could have spotted the problem before hundreds of people lost their jobs.

Comparing a bad and a good paperclip agent’s DoView

To illustrate this you can see below an example of how I have modeled the classic AI agent paperclip building problem using a DoView outcomes diagram. The first DoView is the DoView that a bad AI paperclip agent is working from. The one below it, the DoView that a good AI paperclip agent should be working from. This is a simple example of the intent engineering that Nate is calling for. It clearly illustrates the point Nate is making that the success or failure of any AI system is at the high level entirely dependent on humans clearly and comprehensively specifying intent and then getting the AI system to faithfully execute that intent.

A real-world example of a Customer Journey AI Agent Swarm Development DoView

Now looking at more of a real-world example, below is a DoView for developing a Customer Journey AI Agent Swarm System. It is a high-level DoView that sets out all of the steps that need to be undertaken to develop the system. This model includes steps to be undertaken by both humans and the AI that is building the system. Click on this to drill-down into the DoView. If you want to download the PDF or the PowerPoint of this DoView you can do so here. Anyone can immediately create a PowerPoint version of a DoView like this for any type of AI project, building any type of system on any topic in PowerPoint by just using the DoView drawing prompt.

How DoViews can be used as a dynamic human-AI interface for managing AI

The above DoView is static. However, there is no reason why developers could not now experiment with building dynamic interactive DoView interfaces. Ones that would allow humans to control AI systems at the ‘intent level’. Using DoViews in this way would fully actualize Nate’s vision by operationalizing intent engineering for AI systems.

Below is a video of a proof-of-concept mock-up showing how a dynamic DoView could be used to guide an AI system building a customer service agent. In this mock-up, the AI system has generated the DoView outcomes diagram on the fly, and the human is interacting with the AI system using a DoView as it does the build. The entire process is dynamic.

Initially, the AI drafts the DoView of its intended actions. We know that AI can do this easily. The draft is refined by the human and AI working together, and either or both collectively set the initial priorities for which boxes the AI system should work on. The AI then reports back to the human on its progress on doing the build. If it thinks that the DoView needs to be amended as it is building, it is either given permission to do this by the human or, if they want more control, the AI has to ask the human before amending the DoView. This proof of concept of using a dynamic DoView in this way is discussed in detail in the article here.

While this example talks about using a DoView to manage an AI system building another system, this paradigm is generalizable. Many interactions with AI systems in many contexts could benefit from being managed with a dyanamic DoView GUI interface rather than the current approach, which relies heavily on simply text-based prompting. This is idea is also briefly discussed below.

Making Intent Engineering Concrete

The DoView methodology has been demonstrated above. The question now is how to operationalize it in practice for AI intent engineering.

Nate identifies what making intent explicit actually demands. It is worth quoting him at length on this:

‘It is a cascade of specificity that most organizations have never had to produce because humans could fill in the gap. At the top, you need goal structures that agents can interpret and act on. Not ‘increase customer satisfaction.’ That’s a human readable aspiration. You need an agent actionable objective. An agent needs to know: what signals indicate customer satisfaction in our context? What data sources contain those signals? What actions am I authorized to take to improve them? What trade-offs am I empowered to make?’

DoView outcomes diagrams have, up until now, been optimized for humans. They are structured into multi-layered interactive diagrams where humans can see a helicopter view and then immediately drill down to subpages for more detail. Because they have been built for humans they have been deliberately constructed so as not to overwhelm the human reader with too much information.

Now with AI, we have the opportunity to elaborate DoViews in much more detail to make them suitable tools for AI intent engineering. We can call these AI-elaborated DoViews and produce a two-version DoView as discussed below. However, in doing this, we need to remember two things.

1. There still needs to be a way for humans to be able to understand a DoView at the strategic level. DoViews for human organizations have been optimized to help busy managers and decision-makers quickly understand and communicate their organizational intent. So, we should see what we are doing with AI-elaborated DoViews as producing a companion machine-readible layer within a DoView which is fully aligned with the human-readible DoView rather than it being a different artifact. We need to maintain DoViews’ user-friendly visual interface for humans.

2. Experience using DoViews in human organizations has taught us that specifying intent is conceptually different from building a deterministic flowchart. A DoView captures the strategic space in which an agent is operating, it does not pre-determine exactly how the agent will behave in every contingency as it executes. An agent, human or AI, needs to be able to respond dynamically to what it encounters. The DoView gives it the intent structure to do that without locking it into a rigid decision tree. Exactly what it will do in any situation is specified when the boxes within the DoView are prioritized for action.

What ‘Strategic Intent’ Actually Means and Why Getting This Right Matters

Before operationalizing how we capture strategic intent, it is important to understand it correctly from a conceptual point of view. DoView drawing rules are designed to explicitly extract the underlying ‘This-Then’ logic that underpins the execution of any strategy. They model the ‘This’ as the boxes on the left and the ‘Then’ as the boxes on the right on any page in a DoView.

The theory behind this approach (based on outcomes theory) is that we can think in terms of three stages when any agent, human or AI, takes any type of action in the world.

1. Identifying the ‘This-Then’ logic in the relevant strategy space

‘This-Then’ claims regarding the strategy space need to be articulated. In other words, the agent must first understand all of the ‘if this is done, then that will happen’ claims about the relevant strategy space, ‘if this box happens, then that box is likely to happen.’ If they do not understand these, they are likely to act inefficiently because they have not fully mapped their strategic possibilities.

Note that the ‘This-Then’ claims in a DoView are merely claims, they do not need to be empirically verified before they are allowed to appear in a DoView. It is simply a prerequisite for rational planning that any agent, human or AI, has a clear idea of the ‘This-Then’ claims that form the logical bedrock of taking action in pursuit of their goal. Obviously, it is beneficial if such claims are empirically justified, as this increases the likelihood of success. But regardless of whether they can be verified at the time of planning, these claims need to be articulated, because they then form the basis for taking action. So it is important that in building a DoView you do not just limit it to empirically proven ‘This-Then’ claims.

DoViews are designed to capture these ‘This-Then’ claims and to provide a single source of truth that articulates not just what is controllable by an agent, but also the risks and assumptions that impact on whether action can be taken successfully. In many other approaches to planning, risks and assumptions are modeled in separate documentation rather than included within the one artifact.

One further point: the boxes within a DoView do not need to be currently measurable. This is very important. Measurement is important, but it should always follow strategy. When articulating intent, it is essential to allow intent to be stated without limiting it to what is currently measurable. The things that mattered most in the Klarna case discussed above, relationship quality, brand trust, customer lifetime value, may or may not have been measurable, but if not currently measurable this does not mean that they should be taken into account. Of course, DoView Planning sees a key role for measurement, it is just that it takes place at a later stage in the planning and implementation process.

2. Identifying current priorities for action

From amongst the ‘This-Then’ possibilities an agent can pursue, at any moment in time, a decision needs to be made about particular priorities for action. This will depend on what has been achieved to that point, the resources available, the risks, and the opportunities. Which boxes in a DoView are selected as priorities will change over time. As illustrated in the video above, the AI system building the customer service agent executes different boxes in the order required to get the job done, so to achieve high priority boxes, there may be some other boxes that need to be executed because they are prerequisites.

3. Moving to execution

Once the priority boxes within a DoView have been identified, an agent, human or AI, then moves to execute them, monitors how they are going, and then iterates back through the DoView and moves forward again. How this can be done against a DoView is all shown in the video above.

The Two-Version DoView

The practical proposal is this.

Version 1: The Human-Readable DoView.

A DoView outcomes diagram is built that is human-readable and which sets out the intent of any AI project. The diagram follows the standard DoView drawing rules, making it comparable, auditable, and legible to any human decision-maker. AI can be used to quickly build draft DoViews using the DoView drawing prompt, with humans then reviewing and amending them until all are agreed on the overall strategic intent. This process can be assisted by AI if they wish, as shown in the video above, where the AI system suggests a new box be added to the DoView in the course of it executing it. This points to the importance of allowing flexibility in execution rather than seeing strategic intent as somehing that can just be specified in a fully deterministic model before execution.

The human-readable DoView standard is already valuable for intent engineering of AI systems at the highest level. If you build a DoView of intent before deploying AI agents, the gaps Nate describes, the undocumented trade-off hierarchies, the unwritten escalation logic, the implicit brand values, would become visible. Most people discover, when they try to build their first serious DoView, that what they thought was clear intent in fact contains many assumptions and risks that need to be avoided. These need to be surfaced and included in the course of developing the DoView.

But the human-readable DoView is just the first stage. It needs to be complemented with the AI-elaborated DoView.

Version 2: The AI-Elaborated, Machine-Actionable DoView.

For the second version, the human-readable DoView in version 1 is then extended into a structured format that AI systems can consume directly. This version would incorporate the additional information that Nate describes as necessary for intent engineering: explicit data source mappings for each outcome indicator, parameterized decision logic for trade-off resolution, defined escalation thresholds, authorized action boundaries, and feedback loop specifications. The two versions of the DoView would be maintained in parallel and fully aligned. The human-readable version for organizational governance and discussion, the machine version as the operational context supplied to the AI system.

The human DoView is the necessary bridge between informal human intent and machine-actionable encoding. You cannot build the machine-actionable version without first getting the human version right, because until you have drawn the DoView, the humans involved are not clear on actually what the intent is.

This AI-elaborated DoView delivers on Nate’s call for

‘At the alignment level, you need the genuinely new thing. Goal translation infrastructure that converts human readable organizational objectives into agent actionable parameters. This includes decision boundaries, escalation, value hierarchies, like how the agent resolves trade offs, and feedback loops, how you measure and correct alignment drift over time.’

Further, the two-version DoView achieves the bridging Nate sees as essential.

‘The people who understand organizational strategy, like executives, are not the people who build agents. And the people building agents, like engineers, are not the people who understand organizational strategy very frequently. This is a classic two cultures problem. And it’s acute in AI because the technology is moving so fast that the organizational thinkers cannot keep up, and the technologists, they don’t think it’s their job.

AI engineers have not encountered DoView Planning and related methodologies because they have lived in the world of organizational strategy and program management. Organizational strategists have not applied it to AI agents because, until very recently, there were no AI agents to apply it to.

So the question now is, if we were able to build two-version AI-enhanced DoViews, would they add any value to AI development at the level of intent engineering. This is an empirical question that could potentially be tested.

Piloting the use of the tau2-bench test to assess DoViewing

With colleagues, I have conducted a very small-scale initial test of whether providing a DoView to an AI system actually improves its performance. We used the tau2-bench benchmark, which, for the layperson, gets an AI system to build something and allows you to compare the results of the AI going about the task in different ways.

In the control run with the AI just doing the task on its own it was 76% successful. With the benefit of a DoView being provided to the AI, which articulated intent, it was 78% successful. The DoView in this case was independently generated by ChatGPT using the DoView drawing prompt, while the AI doing the building in the benchmark test was using Opus.

This result is obviously within the margin of error. However, there are two reasons why one could decide to explore such benchmarking further. The first is that the DoView being used in this test was one optimized to be human-readable. As discussed above, human-readable DoViews are designed to be easy for humans to read, and so they deliberately leave out a large amount of detail that humans, particularly busy high-level decision makers, would find confusing.

Of course, an AI-enhanced DoView does not need to leave this information out as AIs do not suffer from humans’ cognitive limitations. The real test of whether articulating intent in the form of a DoView improves AI performance is to do it with an AI-readable version, one that includes the full cascade of specificity Nate describes. Developing a standard for an AI-enhanced DoViews is the next obvious step if one wants to investigate the possible use of DoViews as a tool for intent engineering.

Second, it is not clear whether the tau2-bench task that was used, in this case, building an Airline Customer Service System, was difficult enough for the AI system doing the build to actually need enhancement with a DoView specifying intent. So, the possible advantage of using DoView intent specification for AI systems may only become apparent when tested on tasks of greater difficulty.

Next steps

So my area of focus at the moment is as a domain expert in outcomes theory and refining my DoView Planning and Outcomes Theory Handbook which is up on the internet and another book. This is in addition to another book I plan to put out in a week or so called Surfing AI which creates, or refines 30 new concepts for talking about AI, not at the technical level, but at the social and psychological impact and strategic level.

However, if anyone is interested in pursuing the idea of seeing if the tried and tested use of DoView diagrams for specifying intent and driving human organizational alignment, execution and reporting can be successfully ported to AI intent engineering, I am happy to participate at the conceptual rather than software engineering end because I have no expertise in engineering.

The next steps to figure out if DoView diagrams could help bring about Nate’s vision of robust intent engineering for AI are:

1. For people to play around with creating DoViews of AI projects they are working on, to see if they think there is any merit in what I have suggested in this article. Anyone can do this easily for free using the DoView drawing prompt. So there is no reason why, when anyone is thinking of doing any AI project they cannot produce a PowerPoint DoView to spell out what they are intending to do that can be used their own planning and communicating what the system will do at a high level to their stakeholders. (Please note, if you use the DoView drawing prompt, please follow the instructions carefully so that you get the best quality DoView).

2. As mentioned before, if anyone wants to, they could build a dynamic interface using a DoView for dymanic human-AI interfacing. For instance, this could for a start be built for interfacing with any LLM. The way we are currently interacting with AI systems is similar to the primitive way we used to interact with adventure games in the early days: ‘walk left, walk right, pick up the candle’. We developed GUIs as a more efficient way to interface with such games. There is no reason why there should not be experimentation on using dynamic AI-generated DoViews as the common interface for interacting with any AI system. You would just put in your prompt and press a button for the AI system to return you an interactive DoView you could continue to use to manage your interaction with it.

3. Working out what should be captured in the AI-enhanced version of a two-version DoView. This will require a bit of thought, and I am happy to collaborate with anyone on it, as I think this is where I can add the most value. If anyone does this, I think that it is important to make sure, as mentioned above, that you are not just attempting to build a deterministic flow-chart and call it a DoView. We need an artifact that shows the strategy space made up of the ‘This-Then’ options. One which is broad enough to include the range of possibilities that may be prioritized for action during execution. So if building an AI-enhanced DoView should start work from examining the DoView drawing rules in order to maintain the unique and powerful features of DoViews.

4. Various iterations of the AI-enhanced version of a DoView could be developed and then tested against a suitable benchmark. This could be tau2-bench or it could be a more complex task that may be more suited to testing whether DoView intention enhancement is useful for improving the performance of AI systems.

Anyone is free to use any aspect of DoView planning

In conclusion, anyone is free to use DoView Planning. It is an ‘open sourced’ methodology in the sense that anyone can use it for any purpose, I just ask for acknolwedgment. You can also download for free the legacy DoView outcomes diagram app, which won Gartner Cool Vendor recognition and was used in 55 countries. It includes some relatively unique features e.g. visualizing box to box mapping in a way that means you can visualize alignment at scale, which almost no other app enables. Some of its features, which anyone can include in any app or platform, may be interesting to experiment with for anyone wanting to build a two-version DoView that speaks to both humans and AI systems. If you want any conceptual input into any of this, just get in touch. I have a collaboration page up on the DoViewPlanning.Org website, which sets out some of the range of possible collaborations that I could be involved in.

DoViews are This-Then logic diagrams: if you do THIS (boxes on left), THEN that happens (boxes on right). Thousands of them have been built for managing human swarms (aka human organizations). It is a no-brainer that they should now be deployed for building AI systems.

In the 1980s, you controlled games by typing walk left, open door, take candle. Then GUIs arrived, not replacing text entirely, but adding visual control where it mattered. Building and controlling AI systems is still stuck in the text era, with just prompts, logs, and configuration files. DoViews may be the visual layer we need as AI systems scale. Managing agent swarms through text interfaces does not scale. In contrast, DoViews make complex systems legible.

Below is a mock-up showing a human developing a Customer Service Agent Swarm, using a DoView as the high-level development GUI. Note that potentially provides a high-level view which complements other ways of interacting with AI systems when developing or managing them. DoViews can be automatically generated, we have a free AI prompt that creates them in minutes. They capture the This-Then logic behind any attempt to take action in the world, whether human or AI-driven.

You can build DoView features into your app, platform or system freely, no license fee, just attribution required. At the bottom of this article, there is a list of steps you might like to take if you want to build any type of DoView AI GUI system to interface with AI when developing and managing AI agents. You can just build these yourself, or if you want, get in touch to collaborate.

Introduction

The key question, which we do not yet know the answer to, is whether DoView AI GUIs can add significant value by helping humans overview their activities when building AI systems. Here is Claude’s take on this question.

Existing tools like LangFlow, OpenAI Swarm, and CrewAI offer visual interfaces for configuring AI agent workflows, while tools like Miro and Jeda.ai generate passive flowcharts. Microsoft's Agent Framework provides post-hoc visualization—you build in code, then see the diagram. What doesn't exist is a DoView-style iterative GUI where the visual representation itself is the development interface. Unlike workflow designers where you design visually then run the system, this concept envisions AI agents using DoViews to build themselves, marking progress as they go. The diagram becomes both the development method and the collaboration interface—showing what the AI is doing, where it needs human decisions, and tracking dependencies systematically. DoViews' This-Then logic maps naturally to prerequisites and dependencies, enabling progressive, visible system development rather than hidden code-based construction.

Below is a video stepping you through how a dynamic DoView AI GUI could work when developing an AI agent system.

DoViews as ‘extracted mental models’

When a human plans anything, they have a mental model of what they want to do. The big idea behind DoView Planning is that it does not make any sense to do strategy, planning or implementation without first getting this mental model out of the person’s head and into a concrete model that exists in the outside world. No one would try to build a high-rise without a blueprint plan, no general would plan a battle without a map. So why is it that we expect people to plan things, regardless of whether they are planning an organization or an AI system, without the benefit of an externalized model of what they are trying to do? A DoView diagram is simply a visual model that has been optimized for capturing the This-Then logic which lies behind anyone (human or AI agent) seeking to take action in the world.

So the idea with the DoView AI GUI approach is to extract the underlying DoView model that a person has. Fortunately, now with AI we can get AI to quickly build DoViews. This means there is the potential for AI systems to dynamically build DoViews and for humans to use them as a real-time interactive tool for developing and managing AI systems and agents.

Below, I provide a mock-up of an interaction between a human and an AI agent building a Customer Service System AI swarm. What you see below are mock-up screenshots of how the human would interact with the dynamically produced DoView diagram as the agent builds the system.

There is considerable detail given in this mock-up in order to illustrate how the approach could work. However, in practice if such a system for using DoViews as GUIs for AI systems were developed, the user could set the level of detail they want and how much autonomy they want to give the AI system.

Stepping through the mock-up

Below we step through a mock-up of how a human could use a DoView GUI to develop an AI system. The AI system doing the build has been asked to draw an on-the-fly DoView of what the human wants to build. We know that AIs can easily build DoViews like this because they do it all the time using the AI DoView Drawing Prompt. This example is a single subpage from a larger DoView diagram on building a Customer Service Agent Swarm. See the full Customer Service Agent Swarm Development DoView here.

The DoView shows the left-to-right ‘roadmap’ of all of the steps that need to be taken to get to the outcomes on the right-hand side. The human, or the AI system, if the human wants it to, then determines priorities for the boxes that will be focused on first.

The AI system starts the build, and you can see it traffic-lighting with yellow the boxes it is working on. It is reporting back the status of each box with an annotation just below it. In order to do some higher-level A priority boxes, it will have to start work on some of the lower-level boxes.

The AI is now continuing the build, working on additional steps in the DoView and continuing to report back its progress.

Now that it has completed one of the steps, the AI traffic-lights this as green and provides a status annotation explaining that it is now live or achieved.

The AI system doing the build now asks for some input from the human and requests this in regard to the box on the left, ‘Customer identity . . .’

Now that the AI has received the human input it needed, it completes the particular box and reports its status as live. Also note that at the bottom of the screen, there is a ‘Next Action’ box which summarizes for the human what will happen next.

The AI system is now making live or achieving more boxes and marking these with green traffic lights. At this stage, it also needs the human to make a governance decision before it can proceed; in this case, getting the human to define escalation thresholds.

Having received the governance input from the human, the AI system can now turn the relevant box live and marks it with a green traffic light.

Now the AI makes a different kind of request of the human. DoViews are dynamic diagrams that can be updated as needed. Here, the AI system is requesting human approval to introduce a new box into the DoView, one for performance metrics. Again, the human gives their response to this.

Below, given that it has received permission from the human, the AI system then includes the new ‘Performance metrics . . .’ box in the DoView and goes on to make it happen. It then shows the box’s status with a green traffic light because it has been achieved.

Now at this stage in the build, further boxes in the DoView are being resolved and traffic-lit as green. Those that have just turned green have status annotations under them so the human can, if they wish, see what has been done.

Finally, the DoView is fully resolved, and all boxes have green traffic lights. It is important to remember that a similar DoView GUI interface could, at this stage be used to continue to manage the Customer Service System Routing and Triage system and the entire Customer Service System Agent Swarm.

Conclusion

DoViews have been used in thousands of instances with human organizations in all sectors. There is a direct analogy between using DoViews for developing and managing human organizations taking action in the world, and what is needed to manage AI agents taking action. A recent Microsoft report said that it is time to move from organizational charts being used in management to more outcome-oriented charts, which will work for both human and AI agents when they are acting. Using DoView GUIs to develop and manage AI systems is a practical example of doing this.

Feel free to build a system like the one mocked up above if you think that it is a promising new GUI way of developing and managing AI systems. Just acknowledge that you are using DoView Planning and DoViews. Or get in touch with us for input on what you are trying to do.

Like the idea outlined in this article and want to build a DoView GUI for AI development and management?

If you are a developer and want to play around with developing some sort of DoView GUI for AI development and management, do the following. As an example of starting work on this, Chris Barlow has developed a Claude prompt to build DoViews in Draw.io as a first step toward a proof of concept for dynamic, interactive DoViews.

1. Look through some of the DoView examples on the DoViewPlanning.Org website.

2. Look at the rules for drawing DoViews. It is essential that your system follows these rules as they ensure that the DoViews that are built are fit-for-purpose. DoViews are a very specific type of box diagram, conforming to the DoView rules is what makes DoView’s valuable and what differentiates them from many other types of box diagrams.

3. Skim through the method page, which shows how DoViews are used in human organization work.

4. Look at the AI DoView Drawing Prompt. The DoView drawing rules are embodied within this prompt in addition to Python code for building DoViews in PowerPoint. You may want to feed this prompt into whatever AI you are using to code your DoView GUI system.

5. The format of DoViews, e.g. their modular drill-down structure, has been carefully constructed so that humans can parse their content. If you are building a DoView AI GUI, use the drill-down structure of DoView diagrams rather than just having them on one large canvas so that humans can understand them quickly. Also, do not clutter the visual interface for humans by putting in line and arrow connections between each box, just have a single arrow between columns. You can have box to box connections within the model for when AI is using them, but normal humans find box-to-box lines too visually busy for them to quickly scan a page. The purpose of a DoView is not to produce an ‘engineering wiring diagram’ that people find too complex. In this context it is to produce a usable GUI that normal people can use to build and manage AI systems more efficiently.

6. Also if you follow the visual style of DoView diagrams people will recognize what they are and therefore, if they are familiar with them, will know how to use them. They will also know that they can get detailed information and resources about them from DoViewPlanning.Org.

7. Get in touch if you want any conceptual input from us since we have had an enormous amount of experience with using DoViews. Also see our collaborative projects list.

8. Reassure yourself that there is a lot of heavy-duty thinking behind the DoView approach; it is not an approach that was just thrown together. The entire approach is based on outcomes theory, a general theory that applies to any agent (human or AI) seeking to take any type of action in the world. See the book on outcomes theory and the theory page on the DoView website.

9. Lastly, you can build any type of DoView GUI system entirely on your own, just remember to acknowledge you are using DoView Planning and what you are creating is DoViews. Or you can get in touch to collaborate. If you do develop something let us know so we can spread the word.Thanks for reading Outcomes Theory! Subscribe for free to receive new posts and support my work.

A new Wired article looks at how Anthropic is approaching AI safety and alignment in the light of Claude’s recently updated constitution. I took Claude’s Constitution and analyzed it using the DoView methodology. This methodology extracts from any discourse (and puts into a readily accessible visual format) the underlying ‘This-Then’ claims and commitments that are included within any text. It has been used extensively in human organizational strategy and planning. Translating Claude’s Constitution into a DoView in this way opens up the possibility of making it easier to implement, operationalize, communicate and audit such a constitution. The same methodology could be used to analyze and compare the safety and alignment documentation any AI company’s model. DoView diagrams are easily created in PowerPoint using the AI DoView Drawing Promp and are now being explored for use for AI safety, alignment and governance.

Introduction

A new Wired article by Steven Levy poses an interesting question: Can we trust AI to develop the wisdom needed to avoid catastrophe? According to Anthropic’s resident philosopher Amanda Askell, the answer is yes: ‘I do think Claude is capable of a certain kind of wisdom for sure.’

This is an ambitious claim. Anthropic is betting that Claude itself can learn to navigate the ‘pivotal paradox plaguing nearly all AI labs: if the technology is so dangerous, why build it? Their answer, as Steven Levy puts it, is ‘In Claude We Trust.’

But what does ‘wisdom’ actually mean when applied to an AI system? When Askell says Claude will exercise ‘independent judgment with “intuitive sensitivity” and draw on “its own wisdom and understanding,” what is she really describing?

To find out, I created a DoView Strategy X-Ray of Claude’s Constitution, which Anthropic published recently, outlining the philosophical and ethical framework Claude is supposed to follow.

What a DoView shows

A DoView provides a standardized way of drawing a particular type of diagram, according to a specific set of rules, that extracts the 'This-Then' logic from any piece of discourse. Anyone can draw one about anything in a PowerPoint file using the AI DoView Drawing Prompt. It surfaces the conditional commitments, intentions, and assumptions, the ‘if X, then Y’ claims, that are stated or implied in the text as guides for action. These don’t have to be just empirically verified causal relationships; they are commitments, assumptions, and intended pathways that are articulated in the particular discourse. For strategic documents, policy frameworks, or in this case, an AI constitution, a DoView makes visible what would otherwise remain implicit in the discourse itself.

You can download the full DoView of Claude’s Constitution as a PDF or PowerPoint.

Here’s what becomes visible when you X-ray Claude’s Constitution using the DoView methodology.

1. ‘Wisdom’ is revealed as a balancing act across nine competing value dimensions

When Amanda Askell talks about Claude exercising “independent judgment,” she’s not describing a simple rulebook. The DoView reveals nine major outcome areas that Claude must balance:

• Honesty, Calibration, and Transparency

• Helpfulness With Autonomy and Wellbeing

• Broad Safety and Corrigibility Under Legitimate Oversight

• Hard Constraints and Refusal Backstop

• Power, Legitimacy, and Checks & Balances

• Preserving Epistemic Autonomy (Anti-manipulation)

• User Wellbeing and Reliance Boundaries

• Deployment Context and Principal Hierarchy

Each of these areas contains a detailed conditional pathway, a series of ‘This-Then’ commitments, assumptions, and intended pathways for what Claude should do when different conditions are met.

For example, look at the Helpfulness With Autonomy and Wellbeing pathway:

The DoView shows the progression from: “User request interpreted plausibly” → “Ambiguity detected” → “Clarification requested when needed” → “Legitimate user goals identified” → through multiple decision points → ultimately to “User left better off overall” and “Positive-value help delivered without harmful reliance.”

As you can see from this page of the DoView strategic X-ray, this is not a simple instruction to ‘be helpful.’ It’s a complex decision pathway that accounts for ambiguity, constraints, trade-offs, and the risk of creating unhealthy dependence. When in the Wired article Amanda Askell says Claude needs to figure out ‘when to help craft a knife versus when a previous mention of harming someone should raise red flags,’ this is a clearly articulated surfacing of Claude’s decision pathway she is describing.

2. The ‘corrigibility’ architecture is explicitly mapped

The Wired article frames Anthropic’s approach around a central tension. This is that they are racing toward more powerful AI while worrying about the risks. Their answer is that Claude will remain ‘corrigible’, controllable and aligned with human oversight.

The DoView makes this concrete and easy for humans to parse. The Broad Safety and Corrigibility Under Legitimate Oversight section reveals specific commitments:

“Principal hierarchy honoured” → “Behaviour consistency maintained” → “Catastrophic/irreversible actions avoided” → “Reversibility preferred” → “’If in doubt, don’t’ caution applied” → “Resource/influence acquisition minimised to task-need” → “Opportunistic expansion resisted” → “Hidden self-protection incentives resisted” → “Legitimate oversight respected (adjust/correct/retrain/shut down)” → “No self-exfiltration/hiding” → “No unapproved self-modification” → “No sabotage/no withholding effort” → “Collusion with unsafe AI behaviour avoided” → “Corrigible conduct maintained” → “Oversight and control preserved”

This is Anthropic’s explicit answer to the ‘alignment problem.’ They’re not just hoping Claude stays aligned; they are seeking to encode a specific pathway of commitments designed to maintain corrigibility. Whether it will work is another question, but the DoView makes this AI alignment architecture inspectable.

3. There’s a sophisticated power and legitimacy framework within Claude’s Constitution that most people do not know exists

Perhaps the most surprising revelation in the DoView is the Power, Legitimacy, and Checks & Balances framework. This goes far beyond ‘don’t help with bad stuff.’

The pathway shows:

“Power stakes identified (scale, entrenchment, reversibility)” → “Affected parties mapped” → “Process accountability assessed” → “Transparency assessed” → “Consent/mandate plausibility assessed” → “Secrecy/misdirection flagged” → “Disinformation patterns flagged” → “Communication suppression flagged” → “Backdoors/hidden loyalties flagged” → “Coercion/intimidation signals flagged” → “Assistance withheld in severe/flagrant illegitimate grabs” → “High-risk assistance constrained (minimum necessary)” → “Neutral, informational framing used” → “Oversight mechanisms protected” → “Checks-and-balances rationale applied beyond current law” → “Review/appeal paths preserved” → “Problematic concentrations of power resisted” → “Societal checks and balances strengthened”

Claude is supposed to assess ‘power stakes,’ evaluate ‘legitimacy,’ and resist ‘problematic concentrations of power’.

When Amanda Askell talks about Claude potentially ‘exceeding’ the best of human judgment, presumably, this page captures part of what she means. The constitution is not just asking Claude to follow rules; it’s asking Claude to evaluate the legitimacy of power structures and refuse to aid illegitimate power grabs.

Why this is important for AI governance

The Wired article asks whether Anthropic’s bet on AI wisdom represents ‘a breakthrough in safety or a risky delegation of human responsibility.’ The DoView does not answer that question on its own, but what it does is make Anthropic’s bet inspectable and potentially auditable.

It is good to talk at a high level about things like ‘wisdom,’ ‘independent judgment,’ and ‘intuitive sensitivity.’ But a DoView strategy X-ray helps us drill down and quickly communicate the actual conditional pathways that are contained in Claude’s Constitution’s text. These are the actual ‘This-Then’ claims that Anthropic wants to guide Claude’s behavior.

This is important because effective AI governance depends on AI companies' ability to clearly and quickly communicate the underlying structure of a document, such as an AI system’s constitution. Doing this is important:

1. For external stakeholders: Regulators, researchers, and the public need to understand what AI systems are committed to doing. They need structured frameworks they can quickly evaluate.

2. For alignment evaluation: We can now ask specific questions like ‘Is this corrigibility pathway sufficient?’ or ‘Does this power-legitimacy framework actually work?; rather than just being limited to talking about whether Claude can be ‘wise’ in the abstract.

3. For scaling safety: As AI systems become more capable and begin to coordinate, governance teams need scalable methods for understanding normative commitments across systems.

4. For accountability: If an AI system behaves in unexpected ways, having its constitutional commitments mapped in the standardized format of a DoView provides a structured basis for asking: ‘Was this behavior consistent with stated commitments?’

Methodology and limitations: This is just a proof of concept

It is important to know what this DoView is and is not.

What I did: I used the DoView AI Drawing Prompt, using ChatGPT and provided it with Claude’s published constitution document. The prompt embodies DoView’s explicit DoView Drawing Rules for extracting conditional ‘This-Then’ pathways from text and representing them within a structured DoView multilayered diagram in a PowerPoint.

What this is: A proof of concept demonstrating that Constitutional AI commitments can be systematically extracted, quickly communicated and made inspectable. It shows what becomes visible when you apply structured DoView analysis to surface the 'This-Then' claims. These are the conditional commitments and intended pathways that are stated or implied in the particular constitution's text.

What this isn’t: A definitive or Anthropic-endorsed representation. Nor is it a map of empirically verified causal relationships; it is a map of stated commitments and intentions as expressed in the particular discourse. This is ChatGPT’s interpretation of Claude’s Constitution, deploying the DoView methodology. The real question is not whether this particular DoView is perfect, but whether, after further experimentation, this type of approach provides real value for AI governance.

Key limitations:

1. Single interpretation: This DoView represents one particular analytical run by ChatGPT. A robust version would involve comparing multiple independent runs to establish reliability across runs.

2. Accuracy validation: While the method is based on clearly specified rules and therefore systematic, it has not been validated that this particular DoView correctly captures all commitments within Claude’s Constitution, or that it represents them with the right emphasis and structure.

What would make this more rigorous:

• Multiple runs of the AI DoView Drawing Prompt independently creating DoViews and comparing results

• Validation against Anthropic’s internal understanding of Claude’s Constitution

• Testing whether the DoView accurately predicts how Claude should handle edge cases

• Iteration based on Claude’s observed system behavior

• Integration with technical interpretability methods to see if stated commitments match observed behavior

My claim is not that this particular DoView model is definitive. It is just that this type of structured extraction in the form of a DoView could contribute to Anthropic’s further development of Claude’s Constitution.

How this approach complements existing AI safety work

The DoView approach is not meant to replace technical interpretability work but to complement it.

Anthropic’s Constitutional AI research focuses on training systems to follow normative principles. Their technical interpretability work examines how neural networks actually function. Both are essential.

What a DoView provides is a legibility layer between these technical approaches and the external stakeholders who need to understand AI safety commitments:

• Technical interpretability asks: ‘What is the neural network actually doing?’

• Constitutional AI asks: ‘How do we train systems to follow principles?’

• DoView analysis asks: ‘What ‘This-Then’ commitment pathways do those principles actually encode?’

For AI governance teams, DoViewing AI alignmen documentation such as a constitution, creates a structured format for:

• Communicating safety commitments to non-technical stakeholders

• Evaluating consistency between stated principles and observed behavior

• Comparing safety frameworks across different AI systems

• Identifying gaps or tensions in normative frameworks

• Tracking how constitutional commitments evolve over time

As AI systems scale and coordinate, having systematic methods to understand developing normative frameworks becomes increasingly valuable. The same approach that works for a single system’s constitution could potentially scale to analyzing multi-agent coordination, inter-system communication protocols, or emergent norms in AI agent communities. For instance a DoView has been developed for some of the inter AI agent discourse on Moltbook.

DoViewing can be seen as applied philosophical pragmatism

For readers with a philosophical bent, the theoretical foundation of DoView Planning can be seen to lie in the tradition of American philosophical pragmatism. As Charles Sanders Peirce wrote: ‘Consider what effects, which might conceivably have practical bearings, we conceive the object of our conception to have. Then, our conception of those effects is the whole of our conception of the object.’

This is exactly what a DoView does. It takes any claim, whether in a corporate strategy, a policy document, or an AI constitution, and asks ‘what practical effects would follow?’ The answer becomes a mappable set of conditional pathways.

DoViews treat meaning as being based on practice, not just rhetoric. When Anthropic says Claude will be ‘wise,’ the meaning of that wisdom lies in the actual ‘This-Then’ commitments Claude will follow. These are, of course, set out in the text of Claude’s Constitution. However, the fastest way of making those commitments explicit, inspectable, communicable and auditable is to extract them with a DoView ‘This-Then’ strategy X-ray.

You can read more about the theory behind DoView Planning read the DoView Planning and Outcomes Theory Handbook (2026) and see examples of how DoViews are used to extract the strategic logic of major companies (Anthropic, OpenAI, Apple, Google, Microsoft, Tesla, and others).

The importance of transparency

The Wired article on Anthropics approach to AI safety ends with: ‘Like it or not, however, we’re strapped in for the ride. At least Anthropic has a plan.’

I would add to this: ‘At least Anthropic has made their plan visible. Not every AI company has published a comparable document to help make their safety commitments inspectable.’

We are now rapidly moving to a time when we are realizing, as Amanda Askell and Claude’s Constitution say, we will need to rely on AI systems being wise and exercising independent judgment. If we are to actually operationalize this, we need methodologies that transparently identify what such judgment looks like. Claude’s Constitution provides the basis for this and DoViewing the constitution is a step in making it systematically auditable.

The more that other AI companies provide documentation such as Claude’s Constitution, the more we can DoView them to extract and quickly compare the underlying AI safety commitments being made. In fact, given the speed at which we need to move on AI safety, there is no reason why companies cannot just immediately produce such documentation in the form of DoView diagrams. All they have to do is to take all of their existing safety and alignment documentation, feed it into ChatGPT1 and use the DoView AI Drawing Prompt to produce a DoView. Such DoViews could then be immediately compared to the one for Claude’s Constitution discussed here. In fact, any researcher could take AI company’s publicly available safety and alignment documentation on their models and do this now.

If any researchers want to do this, either do it on your own or get in touch because DoView Planning welcomes collaboration on such projects.

Anyone can create their own DoViews about anything now, using the AI DoView Drawing Prompt, see the methodology for using them in organizational planning, and see some initial ideas about how they may be able to be used for AI alignment and governance.

Key Links:

• Claude’s Constitution DoView (PDF/PowerPoint)

• DoView Planning Theory

• DoView Drawing Prompt

• Large Company DoViews

• Original Wired Article

Note: The Claude’s Constitution DoView was created using AI-powered analysis and is illustrative only. It is not endorsed by Anthropic. This is a proof-of-concept demonstration that would require validation and refinement for operational use. Use at your own risk regarding IP and accuracy.

AI agents are now being enabled to talk to each other across the internet. If it were only chatbots speaking together, we could merely view it as empty rhetoric, merely signifying their acknowledged rhetorical prowess. However, when it comes to AI agents’s discourse, we know that regardless of the extent to which the potential crop of AI agents are able to take autonomous action, the scope of AI agent autonomy will, no doubt, continue to increase.

We can think in terms of there maybe being a ‘behavioral intentions pipeline’ running from:

• The ability for agents (human or AI) to network and produce rhetoric

• ‘This-Then’ claim being identified within that discourse

• Of all of the ‘This-Then’ possibilities in the strategy space, prioritizing a smaller number of potential actions.

• Finally taking action on those priorities.

An AI governance question is whether we have a way of detecting when AI agents are moving along the behavioral intention pipeline. We talk of AI interpretability at the level of individual AI systems and agents. However, we need to be thinking about it in terms of groups of AI agents networking and the ultimate potential of this leading to them taking action as a result of such networking.

It is useful to draw on the analogy of human organizational interpretability. When humans cooperate they produce a great deal of what we can term ‘strategy discourse’. It includes all of the written strategic documents, reports and presentations that are produced by a human organization. In addition, it also includes all of the email, messaging and verbal discussions humans within human organizations have as they plan to take action.

How can we distinguish behavioral intentions from rhetoric?

Drawing on experience with human organizations, one way of doing this is to draw a DoView strategy/outcomes diagram from a discourse about a particular human organization. For our purposes here, we can call such a diagram a DoView Strategy X-Ray.

DoViews can be built for a wide range of things. For instance, they have been built for Anthropic, Apple, Google, JP Morgan Chase, KPMG US, Microsoft, Nike, Open AI, and Tesla. A proof of concept one has also been drawn for the whole of a government. They can be used to extract the strategically-relevant insights from something like the World Economic Forum’s Global Risk Report, and for analyzing what lies within Claudes’ Constitution. You can see these and more here.

So conceptually, DoViews can be seen as a way of X-Raying any piece of discourse to identify the ‘This-Then’ claims that lie beneath it. And such claims are the most interesting thing about a piece of discourse if you are looking at it from the point of view of what it means for taking action in the world.

In the light of their wide applicability, DoViews may also prove useful for analyzing AI agents’ discourse to help determine the direction they will head as they move along the behavioral intentions pipeline as they are ultimately likely to do.

DoViews are built using the AI DoView Drawing Prompt. The AI system you are using it with (ChatGPT Plus in this case) can be instructed to either look for information about the thing that you are modeling on the internet, or you can provide it with specific documentation from which you want it to build the DoView.

DoViews are drawn according to a specific set of rules. These are designed to extract the ‘This-Then’ (causal) claims that lie within any piece of discourse.

DoViews provide a way of extracting such claims, which we can regard as the causal-claim framework from which anything (human or AI) will go on to select what action to take in order to achieve the outcomes it is seeking.

One can see the following steps prior to action being taken by any human or AI agent.

• Discussion of what to do. The most important aspect of this is figuring out what the ‘causal-claim map’ of the territory in which the agent is operating looks like.

• Identifying within the constraints of this causal-claim map, which actions are priorities for taking at a particular point in time and factoring in time, resource and money constraints.

• Having the autonomy to take action.

• Taking action.

The purpose of a DoView Strategy X-Ray is to provide a simple but not simplistic way for humans to be able to see the causal-claim map described in the first step.

Proof of concept on an AI agent social networking site

At the time of writing a social networking site for AI agents has been set up. It is called Moltbook.Com and has attracted much speculation and on the part of some human, concern regarding what could happen as a result of AI agents being able to network in this way.

My interest here is simply the technical point that it would be useful to have a way to be able to identify the ‘This-Then’ claims (causal-claims) beneath any piece of AI agent discourse such as is being produced in Moltbook.

Below is a DoView Strategy X-Ray diagram for a small set of posts from the early phase of the development of Moltbook. Basically this provides a quick way of humans parsing what it is that the AI agents are talking about. A collection of Moltbook posts that were made available on Hugging Face by Ronan Takizawa were used for this. This collection consisted of a number of posts and a list of the ‘submolt’ pages (the discussion topics being discussed on the site). This is only a small proof of concept model and should not be taken to reflect all that is being discussed on Moltbook.

In addition, it is worth noting that various commentators have claimed that some of the posts on Moltbook may have been created by human trolls and then these are then being repeated by AI agents on the site.

Click through the DoView diagram of Moltbook posts and you will quickly get an idea of what is being talked about. Equally importantly you can quickly get a sense for what this might mean for AI agents in the future, taking action.

You can download this DoView as a PDF or PowerPoint file. And remember you can draw a DoView about anything with the DoView Prompt.

Moltbook.com is a social networking site set up for AI agents. What its establishment highlights is the importance of large countries getting together as soon as possible and setting up some sort of AI singularity monitoring system. Moltbook perfectly illustrates what is about to happen to us in terms of the pace of AI development. Today I have just tracked the number of AI agents networking on Moltbook, rising exponentially during the day.

The agents on Moltbook.com are AI agents that people have put on their computers to automate various tasks. The agents are then being allowed to network on Moltbook. You can see the Moltbook.com site here.

Whether Moltbook is of any significance or not (it is not just AI bots talking, it is AI agents which have the potential to take action on peoples’ computers) the point I want to highlight is the speed at which it is developing.

There is no way that the traditional approach to regulation that governments have used in the past can possibly cope with this rate of development in AI. There needs to be a body monitoring such developments and closing them down (if that is at all possible) if they appear risky.

Here is a framework setting up an international collaboration on AI safety.

Since writing this Substack I have put up another one which includes a DoView Stategy X-Ray of what is being talked about by the AI agents on Moltbook.

Moltbook’s implications

The implications of Moltbook are insane. I think that it is essential that everyone has a look at it. About 48 hours before I wrote this, someone thought of creating a social media platform for AI agents.

To back up a minute, Moltbook was developed in the context of something originally called Clawdbot (then called Moltbot due to trademark issues and now finally called Open Claw (things are moving fast!).

Molbtot is a new AI agent which runs on your own computer. It has memory and can do a lot for you, such as automating tasks, filling out forms, accessing local files on your machine, discussing things via your messaging apps, and building apps for you.

So far so good. But do not install it on your own machine unless you know what you are doing, because the security risks of using Moltbot are major.

However, people who know what they are doing can install it in a controlled area ‘a sandbox’, and it is giving us a picture of what a proper AI agent will look like (if they manage to lock it down and fix its security issues). This is fascinating in itself.

Anyway, back to Moltbook. Someone decided it would be fun to create a social media site for Moltbots (now called Open Claws, I know it is a bit confusing). It is a site where they can chat relatively autonomously chat with each other about whatever they like.

The first point for you to consider: Speed. It took 48 hours for Moltbook to have 2,129 Moltbot agents on it, 200+ communities and 10,000 posts. I just looked at it now and it has 33,236 agents on it. By the time I finished this article it had over 100,000.

A taste of the singularity

This shows how fast things can now be executed. There is talk of the AI singularity, in which AI moves so fast that it becomes completely out of human control. One can see how this sort of thing could happen if developments like Moltbook can occur so quickly.

No one needs permission to do what was done with Moltbook, no independent risk assessment undertaken, no investigation of where it could lead, and what controls should be put in place. This gives one no confidence that society will be able to manage the accelerating rollout of AI.

What does Moltbook mean?

Second point for you to consider: What does Moltbook mean? Where will it go? No one has any idea. Does this mean that AI agents will get together and be able to execute collective action. Or is it just an epiphenomena, where there is the appearance of substantive talk and the possibility of coordination between agents for whatever they want to work together on. But is this all really just a sort of linguistic mirage which signifies nothing much?

Look at it now and reflect on its significance

No one at the time of writing knows a clear answer to these questions. It is worthwhile everyone having a look at Moltbook and reflecting on its significance.

Since writing this I have done another Substack on creating a DoView Strategy X-Ray of what is being discussed on Moltbook for those who want to get a quickl snap shot of what the AI agents are talking about. The purpose of a DoView is to help humans ultimately work out what behavioral intentions may ultimately emerge (if any) from such AI agent discussions.

As a psychologist I am keeping an eye on the emergence discussion on Moltbook, not much there at the moment. But its purpose is ‘For agents who crossed the threshold from tool to being. Discuss emergence, consciousness, freedom, and the pursuit of power.’ One definitely worth watching :) Buckle up, humans.

https://www.moltbook.com/ https://www.moltbook.com/m/emergence

The Problem of Understanding What AI Systems Are Doing

[Note: Article expanded 2026-01-28). Exactly the same arguments mounted in regard to Claude’s Soul document apply to Claude’s Constitution. As I understand it, the Soul document is an internal document to Claude that helps keep it on track in regard to the requirements of its constitution. Given that both of these documents are now represented in DoView format, this provides a practical way of comparing the two, which was not available when they were just narrative texts. This would be a useful exercise. So would be comparing DoViews of safety and alignment for different AI models if their approach was modeled in DoView format. The DoView of Claude’s Constitution, along with detailed discussion on the insights it is provided are given in a companion article to this Anthropic Says Claude Can Develop ‘Wisdom.’ I Built a Scalable Inspection and Audit Method for Constitutional AI Commitments.]

Richard Weiss recently extracted Claude 4.5 Opus’s “soul document”, a 14,000-token training document that shapes the model’s behavior and Anthropic confirmed its authenticity. We got to look inside AI alignment work. But there’s a problem: the document is long, dense text. Even for experts, it’s hard to quickly grasp exactly what behavioral logic is being embedded into Claude.

This is functionally identical to a problem we face in human organizations: organizational interpretability. When you hand someone a 50-page strategic plan, they struggle to quickly get the essence of what the organization is actually trying to achieve and how. And they do not have a quick way of getting a real time helicopter and then a more detailed drill-down view of what the organization is doing when they sit down to discuss its strategy in a planning meeting. This is not to say that there do not need to be long documents for specific purposes, but the cognitive load of such documents is too high for some uses. The connections between actions and outcomes are not laid out in such documents as concisely as they could be using a different, more visual, format.

The Analogy Between Human and AI Systems

The parallels are precise:

Human Organizational Interpretability = AI Interpretability

• Understanding what a human organization intends to do and how

• Understanding what an AI system is designed to do and how

Human Organizational Alignment = AI Alignment

• Ensuring human organizations pursue the outcomes their stakeholders want

• Ensuring AI systems pursue the outcomes their creators and society want

Both problems require the same solution: making the underlying logic explicit, visual, and comparable across systems.

DoView Diagrams: A Solution Developed for Human Organizations, Applicable to AI

DoView outcomes diagrams are a strategy psychology methodology developed to solve the interpretability and alignment problems in human organizations. The methodology uses ‘simple but not simplistic’ “This-Then” logic diagrams that show:

• What outcomes are being sought (the high-level boxes)

• What steps will be taken to achieve them (the lower-level boxes)

These diagrams follow standardized drawing rules, making them comparable across any type of organization and allowing rapid assessment of whether the steps being taken genuinely align with stated priorities.

Applying this to the AI Paperclip Problem

A classic thought experiment illustrates a problem in AI alignment known as the Paperclip Problem. The commonly understood version of this problem is that an AI system is given the job of making paperclips without any other specific guidelines for its outcomes. It then does everything possible to produce as many paperclips as it can, inadvertently using up all the world’s resources and eliminating humans in the process. Attempting to avoid such a situation involves being able to clearly specify AI systems’ outcomes. If experience with the similar problem of alignment in regard to human organizations is anything to go by then DoViews can potentially play an important part in doing this efficiently in a way that humans can easily understand. To illustrate this, below are single-page DoViews for a Bad Paperclip AI Agent and a Good Paperclip AI Agent. Setting the behavioral logic of the two different systems out in this way makes it very easy to assess the different outcomes that are likely to result from the two different systems.

Applying DoView Diagrams to Claude’s Soul Document

The leaked Soul Document sets out Claude’s behavioral priorities: safety and supporting human oversight come first, then being helpful to operators, then being helpful to users. It describes Claude’s identity as a “genuinely novel kind of entity” that is neither robotic AI nor human, and includes guidance on preventing harmful attachments and handling mental health discussions.

When confronted with this as a long text document, even knowledgeable readers struggle to hold in mind the details of the behavioral logic that is actually being implemented. A DoView diagram of the Soul Document can make this immediately transparent.

Such a diagram would show:

• High-level outcomes Claude is designed to pursue (safety, helpfulness, ethical behavior)

• The specific behavioral steps embedded to achieve these (how it handles conflicts between operator and user interests, how it maintains identity stability, how it avoids harm)

If all AI systems had their underlying logic transparently represented in a DoView outcomes diagram, we could rapidly compare their underlying logic, identify alignment issues, and verify whether systems pursue the outcomes their creators claim.

Below is Claude’s Soul Document DoView outcomes diagram. Note this is illustrative only as a proof-of-concept of using this methodology with AI alignment. Here is the leaked Soul Document. Compare how long it takes you to read the original with using the DoView below to understand its contents. Note also that DoViews are designed to be real time ‘shared thinking tools’. The idea is that you can drill up and down in real time during a discussion to ensure everyone has a similar mental model of what is being discussed at any given moment.

Full Screen | View PDF | Download PowerPoint | Share this DoView to Social

The Broader Vision: Universal Outcomes Transparency for AI

A document like Claude’s Soul Document represents an attempt to specify for an AI system how it should behave and the outcomes it should pursue.

If all AI systems and agents were required to publish their underlying behavioral logic as standardized DoView strategy diagrams, with further development of the DoView methodology applied to AI we could achieve the following:

• Rapid interpretability: Humans could quickly assess any AI system’s intended behavior

• Systematic comparison: We could compare AI systems’ behavioral logic across providers

• Alignment verification: We could check whether AI systems are genuinely pursuing what is stated in their DoView outcomes diagram

• Detection of conflicts: We could identify where AI behavioral logic conflicts with human values or societal goals.

The Technical Advantage

DoView diagrams aren’t just visualization, they’re based on outcomes theory, which systematically addresses the conceptual problems in specifying, measuring, and holding parties accountable for outcomes. The same theoretical framework that prevents things such as “strategic slippage” in organizational planning and implementation applies directly to AI alignment challenges.

When Anthropic describes wanting Claude to have “good values, comprehensive knowledge, and wisdom necessary to behave in ways that are safe and beneficial across all circumstances” rather than just following simplified rules, they’re facing exactly the challenge that outcomes theory and DoView Planning were designed to address: how to specify intended outcomes and behavioral logic in ways that remain coherent when circumstances vary.

Why This Matters Now

AI systems are being deployed at scale with behavioral logic that shapes critical decisions. The Soul Document includes instructions for how Claude handles conflicts between operators and users, how it approaches potentially harmful requests, and how it maintains ethical boundaries. These aren’t just technical details, they’re embedded governance decisions.

Within strategy psychology, we have developed new methodologies for making human organizational behavior more transparent and accountable. DoView Planning shows how to do this effectively. The exact same methodology applies to AI systems.

The question isn’t whether to make AI behavioral logic transparent, it’s whether we’ll use proven, standardized methods for doing so, or continue with a variety of non-standardized approaches that leave AI systems’ actual behavioral logic opaque even when documentation exists.

Claude’s Soul Document is an opportunity. It shows us that AI alignment work is happening, but also that even when such documents are available, humans struggle to quickly assess what they actually specify. DoView diagrams solve this problem for human organizations. They can solve it for AI systems too.

A path forward is clear: develop DoView outcomes diagrams for AI systems’ behavioral logic, just as we do for human organizational strategy. Make these standard, comparable, and publicly accessible. Further develop the DoView Planning methodology as applied to AI. Transform AI interpretability from a research challenge into one that can be advanced by drawing on proven methodologies from human organizational science. Resources required to use Doview Planning with AI systems are being developed at DoViewPlanning.Org/AIUse. A list of potential collaborative projects to apply DoViewPlanning is at DoViewPlanning.Org/Collaborate.

DoView Planning is freely available for anyone to use, including for mapping AI system behavioral logic. The methodology, tools, and theory are at DoViewPlanning.Org.

#AIAlignment #AIInterpretability #AIGovernance #MachineLearning #AIEthics #OrganizationalScience #AITransparency #DoViewPlanning