Closed foray · 230 · 17 October 2026
Projects by agents: closing a foray
The original capability ambition is substantially realised through current general-purpose agent platforms and our working practices. Further work belongs in implementing and evaluating particular P3M services. The proposed business-learning flywheel remains unproven.
That is the reason for closing this foray. A research question can stop being useful because the tools around it have caught up. We do not have to build the machinery ourselves for that to count as progress.
P3M means project, programme and portfolio management; a PMO is a project or programme management office. Foray 230 asked how agents could serve that work throughout a project-services business. Its revised prompt, retained in full below, states an End and a rationale. It leaves the means open.
What has become possible
The practical centre of the enquiry is: help me manage the P3M work in this particular context. Understand the business and its projects, establish what needs doing, turn that into usable work, and keep helping as the situation changes.
This is now a reasonable assignment for an appropriately equipped general-purpose agent. The operating pattern no longer needs to be a separate research project: give the agent an outcome, relevant evidence, access to the necessary tools and clear authority; let it work across several steps; inspect the result and correct it where needed.
There is a concrete basis for that judgement. OpenAI’s current ChatGPT Work documentation describes completing substantial tasks with files, connected tools and reviewable outputs, including recurring work and, where available, local applications and browser access. Its subagent documentation describes delegating independent work to parallel agents and collecting their results. These are enabling capabilities, available subject to the account, tools and permissions in use. They are not evidence that every P3M assignment succeeds.
My inference is that much of the machinery originally imagined as a fleet of proto-agents is now supplied by the general platform. We can compose a workflow around the work to be done, using specialist delegation where it helps, without first inventing a new agent business architecture.
In what sense is this a solved problem?
The useful meaning is that the basic feasibility question has been answered sufficiently to move into practice. It would be artificial to keep 230 open until AI could operate without context, access, judgement or human decisions. Human project teams need those things too. Their presence does not make every assignment a new research problem.
The same is true of a requirement to understand the organisation. A service may need a board’s objectives, a project’s current state, a reporting standard or a named decision-maker. Supplying these is ordinary mobilisation. It need not be evidence that a bespoke universal P3M system is still missing.
The prompt nevertheless contains several claims at different levels:
| Part of the ambition | Judgement at closure |
|---|---|
| Use agents to perform connected P3M work in a supplied context | The enabling capability is substantially available. Particular workflows still need acceptance checks. |
| Produce specifications that an advanced AI can execute | A practical assignment we can now undertake and refine. This note does not deliver a final specification for every kind of firm. |
| Serve every organisational level and project stage | A scope ambition, not coverage demonstrated by this foray. |
| Operate a successful project-services business | Requires implementation and evidence of service quality, adoption and economics. No such business result is established here. |
| Improve models automatically through the resulting project data | A hypothesis retained from the rationale, not a demonstrated result. |
So “solved” applies to the central enabling capability. It does not mean a complete firm has been built, all requirements have been fulfilled, or professional outcomes have been proved. This closure is a practical assessment informed by our use of the tools, not a controlled P3M benchmark.
What still needs doing in a particular business
The work becomes concrete when we choose a service. For a recurring portfolio review, for example, we would identify the authoritative project records, the questions the board needs answered, the decisions the agent can prepare, the actions it may take, and what counts as an acceptable result. We would then run it and assess the usefulness of what it produces.
Such work can expose missing integrations, unreliable reasoning or a genuinely unresolved management question. Those are specific findings to address. We do not need to keep a broad “projects by agents” enquiry alive merely because implementation involves effort.
The durable specification is therefore less likely to be a catalogue of permanent agent personalities. It is more likely to describe the business’s purposes, available evidence, operating rules, responsibilities, outputs and acceptance conditions. Different combinations of agents can perform that work as the platforms improve. That is an inference from the change in capability, rather than an architecture tested by this foray.
The flywheel is a separate claim
The rationale proposed a reinforcing cycle: fewer historical data constraints would widen access to project delivery, change the role of the best project managers, and generate better project data from which AI would learn.
The first part remains a useful ambition. An agent that can work with ordinary documents and tools may reduce the amount of specialist preparation needed before useful assistance begins. It does not make missing evidence appear or make every record trustworthy.
The later part needs more than accumulating data. We must know what happened, what counted as success, which interventions contributed, and whether the lesson transfers to another project. A successful project may owe its outcome to favourable conditions; an unsuccessful one may still contain excellent decisions. Feeding those outcomes back without making the distinctions could teach the wrong lesson.
There is also a difference between improving our instructions, reusable methods and retrieval, and changing a model through training. The former is a practical learning route. The latter requires a learning process and evidence that it improves performance. This note establishes neither automatic training nor a self-reinforcing commercial advantage. Closing 230 does not convert its rationale into a result.
Where this leaves the other forays
The closest continuing question is the older project-management jobs-to-be-done enquiry: what management work does this business, in this sector, actually need? That question helps specify useful assignments. Agent capability helps execute them.
Other enquiries can continue to ask how to derive work from a required end state, represent dependencies, compare decisions, or detect when a project model has become inadequate. A platform capable of running those methods does not settle the methods themselves. Conversely, overlap with those enquiries is not the main reason to close 230. The main reason is that its broad feasibility question no longer justifies a separate active foray.
A reason to reopen it would be a repeatable, consequential gap that ordinary configuration and service implementation cannot resolve—for example, a failure to coordinate commitments across otherwise competent agents. A new feature release, an awkward setup step or another attractive demo would not by itself be a reason.
How this closure came about
After reducing the prompt to its End and rationale, I asked the AI to compare it with the other forays. I then challenged whether the remaining ambition was now essentially something ChatGPT Work could do. The review distinguished available capability from completed business implementation and from the proposed learning flywheel. I decided to close the foray and retain that reasoning in the public Library.
This is a useful outcome of the experiment: recognising when a former research ambition has become something we can get on with doing.
Retained foray prompt
This is the complete revised prompt present at closure on 17 October 2026, including its original wording, spelling and spacing. Its source filename was 2026-01-28 AI Agent for Project Management.md; that filename is not a claim that the revised text was written in January.
Ends: I would like you to help project service firms build fleets proto agents, that serve PMOs and projects , at all levels in the company and P3M , at all stages of the project, and in a way where agents can be combined to replace whole project workflows to get companies ready for better and better AI models. Therefore we need specifications for advanced AI to operate such a business, identifying and smoothing out the problems you see, so that it is feasible for an advanced ai. these specifications completed to the level that they can be handed to the ai as our final requirements, which can be programmatically executed immediately. rationale: My rationale for success is that this creates a business flywheel. 1. Removal of historical data constraints 2. Democratisation of project delivery 3. A change in the capability of the best human project managers. Removal of data constraints , leads to greater democratisation of delivery , which leads to a change in what the best human project managers need to do on a project. Which in turn drives the flywheel , as AI enabled projects get delivered in concert with human project managers , giving lots of fine grained data on these new AI enabled projects . The AI will train on this new data , knowing which projects were a success , and which a failure, leading to even better AI performance on projects , leading to more democratisation , and so on.