AI Mechanics · an essay in the front matterThe Manifesto, Re-Examined
Which of the four still holds when the collaborator is not a person?
Four values, written in 2001 for a room with only people in it. Twenty-five years later there is something else at the table. This is what that does to each one.
- Method
- Each value gets the same three questions: what it got right, what AI complicates, and what survives.
- Posture
- An extension informed by experience, not a replacement. The 2001 authors were right about the work in front of them.
- Status
- v0.1. A working draft, published as one. The declaration at the end will be revised by rooms, not by us.
- Length
- About nine minutes.
15Before the four
Four values, written for a room with only people in it
In 2001, seventeen practitioners wrote four values on a page, and the industry spent twenty-five years arguing about them productively. That is a rare outcome for a document. It happened because the Manifesto described what its authors had already found to be true, rather than what they wished were true.
The four values assume a particular shape of work: people, together, building software for customers, learning as they go. They assume capacities that everyone in the room had. Judgment. Memory that carries from one conversation to the next. The ability to notice that something is off before anyone can say why. Accountability that cannot be handed anywhere else.
AI introduces a collaborator with none of those capacities and a great deal of speed. That does not make the Manifesto wrong. It changes how each of the four values lands when you try to use one on a Tuesday afternoon. What follows walks the four in order and asks the same three questions of each.
One disclosure first. This is written from inside the work rather than about it: a consulting practice that builds software with AI agents daily, keeps a decision log, and has receipts for its own mistakes, including several quoted below. It is not research. Where a claim comes from one shop’s experience, it says so.
- Got right
- Every insight that changed how we work arrived through a conversation. None arrived from a process.
- AI complicates
- Process and tools grew in necessity, because the new collaborator lacks the capacities the value assumed.
- What survives
- The ordering, stated more precisely: tools serve humans, and process serves the agent.
- Bureaucracy
- Process aimed at people who did not need it. Costly, and the thing the Manifesto was written against.
- Scaffolding
- Process aimed at an agent that cannot ask, cannot notice, and does not remember you. Cutting it does not free anybody; it just removes the guardrail.
Re-examined
Tools serve humans. Process serves the agent.
The first half of this value has held up better than almost anything else in the document. Go back through two years of decisions and ask where each one came from, and the honest answer is always a room, a call, or somebody saying the quiet thing out loud on a Friday. No tool surfaced any of it. No process generated any of it.
What AI complicates is the second half of the sentence. A human collaborator asks when uncertain, notices when something is off, adapts without being told, and remembers you next week. An agent does none of those by default. So when an agent enters the workflow you have to add, explicitly, what a person supplied for free: gates, written specifications, verification rituals, and context handoffs, because the agent will never think to ask.
There is an uncomfortable implication in that. In the human-only era, process existed to serve people, and any process that did not was overhead. In AI-assisted work, a large share of process exists to compensate for what the agent cannot do. That is a different kind of process with a different justification, and calling both by the same word is how teams end up cutting the wrong one.
What survivesThe ordering holds, with the reason updated. When a team argues about whether a new gate is bureaucratic, the useful question is not how heavy it is. It is who it is for.
2- Got right
- Code that runs beats documents describing code that might. Still the truth standard.
- AI complicates
- The written intent is now the contract the agent acts on, and it is read every single time.
- What survives
- The measure. What changed is that the shortest path to working software runs through more written intent.
Re-examined
A vague specification does not produce a vague result
The truth standard is intact. Our own gates measure a running application: does the route return, does the page render, does the journey complete. When something breaks in production the fix is code, not a document about code. Nothing AI has done softens that.
What AI complicates is the path to the running code, because the specification has been promoted. In human-only work the code was the truth and the document was an explanation written afterward and read rarely. In AI-assisted work the written intent is what the agent acts on, and it is read on every run. A vague specification does not produce a vague result. It produces a confident, complete, plausible result that is wrong in a way nobody notices for two weeks.
That inverts an instinct, and it is worth sitting with. The Manifesto’s authors were reacting against documentation that had stopped serving the software: hundreds of pages, signed off, out of date on delivery. The documentation AI work requires is a different animal. It is short, it is current, and something reads it and acts on it immediately. If nothing reads it, it is the old kind, and it should be deleted.
What survivesWorking software is still how you know. Written intent earns its place by being acted on, not by being filed.
- Got right
- Discovery beats a template. Sixty minutes of listening beats a proposal written from a category assumption.
- AI complicates
- Ownership, data handling, and sign-off became loud questions that goodwill cannot answer.
- What survives
- Collaborate first. Write the boring things down early, so the collaboration can run at the speed AI makes possible.
- A contract that protects you
- Written for the argument you expect to have. It arrives late and reads as distrust, which is usually accurate.
- A contract that protects the work
- Names who owns which component at the moment it is created, so nobody has to reconstruct it from memory nine months later.
Re-examined
Contracts are not the enemy of collaboration. Vague ones are.
Every engagement that has gone well started with discovery rather than a template. That has not changed, and it will not. The value earns its position.
What AI complicates is the ambiguity that collaboration comfortably leaves behind. Three questions that used to be quiet are now loud. Who owns the output when a model produced most of it. Which client data may be pasted into which system. Who signed off, when the analysis assembled itself in forty seconds. None of these are hostile questions, and none of them are answered by goodwill.
The instinct is to read that as a retreat: more contract, less collaboration. It is the opposite, and it is the sharpest thing in this whole re-examination. A client who does not know whether the material we build together is theirs will hold back the exact context that makes the work good. Naming ownership at creation, rather than in a dispute afterward, is what makes the collaboration safe enough to be honest.
What survivesThe priority, with the reason updated. Clarity is not the price of trust. It is the thing that lets trust move quickly.
- Got right
- More alive now than in 2001. A twelve-month roadmap is fiction by month four.
- AI complicates
- The cost of having no plan rose. Unplanned teams used to move slowly; now they sprint in several directions at once.
- What survives
- Responding to change beats following a plan. It does not beat having a discipline.
- The roadmap
- Twelve months, committed. What the Manifesto warned about. We do not write these.
- The tracker
- Short cycle, assumes it will be wrong, cheap to change. Compatible.
- The doctrine
- Not a plan at all: a decision aid for situations nobody predicted. The part that holds when the rest moves.
Re-examined
Without a discipline, fast response is just fast
This value is more alive now than when it was written. Capability roughly doubles and price roughly halves on a cycle short enough that a year-long plan is fiction by the spring. On 12 June 2026 two production models were withdrawn from general access with no notice. Teams that had bet a plan on a specific model lost weeks. Teams that had paced themselves to trust rather than to capability kept shipping.
What AI complicates is the cost of having no plan at all. When generation was slow, an unplanned team simply moved slowly. When generation is instant, an unplanned team sprints in four directions at once and pays for all four. A receipt of our own: one agent left running overnight against a paid API produced a hundred dollars of spend by morning, and nothing anybody had decided to want.
So plans came back, in a specific shape. Not the roadmap. Two smaller things: a short-cycle tracker that assumes it will be wrong, and a doctrine, which is not a plan at all but a way of choosing when the situation is new. Both are cheap to change. Neither predicts a quarter.
What survivesThe value, with a distinction underneath it. In fast conditions, the discipline is the only thing that makes responding coherent instead of merely quick.
Cross-cutting
Read the four together and the same three patterns surface
One. The Manifesto assumed capacities the new collaborator lacks
Each value presumes something about who is in the room. Individuals who can interact. Builders who notice when the software is off. Customers and developers accumulating trust across years. Judgment available to override the plan when reality shifts. The right-side items the Manifesto deprioritized, which is to say process, documentation, contracts and plans, have grown in necessity not as a rejection of the left side but as compensation for capacities the agent does not have. That is how the Manifesto can be entirely correct and still leave a team short.
Two. The human-AI boundary is the new subject
The Manifesto has nothing to say about the boundary, because in 2001 there was no boundary to say anything about. Yet almost every hard decision in AI-assisted work is a boundary decision: what the agent may decide, where a person must touch the work, what the handoff carries, whose name goes on the result. The most useful sentence we have found is that the agent is never accountable. That is less an ethical position than an operational one. Accountability has to land somewhere and it can only land on a person, so the real design question is which person, at which step.
Three. What compounds has changed shape
The Manifesto’s answer to what accumulates was working software. In AI-assisted work the things that compound are the specification, the verification ritual, the doctrine, the tagged ownership of the material, and the shared vocabulary a team uses to argue about the other four. Every one of those is closer to documentation than the Manifesto preferred. Every one of them exists to produce more working software than the Manifesto thought possible. The tension is real, and it resolves in favor of both.
The draft
Five values, because the boundary added a dimension
What follows is a draft, and the version number is the honest part of it. Five rather than four, because the human-AI boundary is not covered by any of the original four and does not fit inside them.
We are uncovering better ways of building and operating systems by doing it with AI and helping others do it. Through this work we have come to value:
- 01Human judgment over AI execution speed
- 02Clear specifications over hopeful prompts
- 03Verification rituals over output volume
- 04Component-level clarity over wholesale ownership transfer
- 05Operationalized belief over imposed framework
That is, while there is value in the items on the right, we value the items on the left more.
Publishing that as settled would break the argument this essay just made. The four values it re-examines earned their standing across twenty-five years of rooms, not in the week they were written. These five have a hundred-odd rooms behind them at most. They will be revised by rooms, not by us. Run them against your own work, and the ones that do not survive contact should be cut rather than defended.
The formal version of the same thinking, written as a doctrine rather than a declaration and tested against five change-tests, is on the home page. If you wanted the argument for why it is shaped that way, you have just read it.