Architecting the AI Coworker
You are not delegating to a model. You are architecting a team.
Twenty-two essays. Five arcs. One operating model for the AI coworker. Written for the people who own the consequences when an agent acts.
Every essay also ships its own kit — the essay, a worksheet, and the LinkedIn carousel deck, all downloadable from the essay itself.
Jurisdiction note. Written for teams operating across the US, UK and EU, with Canadian and global standards used where they are the best example. It is not legal advice. Where law appears, it sharpens the engineering question: what must the system be able to show, contain, reverse, explain or assign to an owner?
How to read it
I’d encourage you to read the series as a whole. It is one argument in twenty-two moves — each essay assumes the one before it, building from what an AI coworker actually is, through the runtime where it acts, to how you architect, govern, and stand behind one. Read end to end, it changes how you see the entire stack, not just the part you own.
If you don’t have time for all of it, start where your responsibility sits and follow the thread outward:
- CEOs & founders · ~45 min: There Is No "It", Stop Delegating. Start Architecting.
- CTOs & engineering leaders · ~72 min: The Reliability Equation, Tools Give Models Hands, Compliance Is Not a PDF
- AI engineers & security architects · ~65 min: The Nine Layers Where Agents Break, The Sentence That Owns the Agent, Show Me the Run
- Buyers & procurement · ~77 min: Demo Is Not Deployment, The Model Card Won't Save You, You Cannot Benchmark a Coworker
- Governance, legal & compliance · ~88 min: What an Agent Remembers, and Cannot Forget, When an Agent Acts, Who Acted?, Compliance Is Not a PDF
- Skeptics · ~62 min: The Coworker Illusion, The Model Cannot Mark Its Own Work, The Dashboard Is Green. The Meaning Is Wrong.
However you start, the whole series is the real argument — these are just doors into it.
The five arcs
See the Object
What is the AI coworker, really?
- 1There Is No "It"Audit asks 'who approved this?' The only honest answer: 'the stack did.'
- 2The Coworker Illusion"The AI did it" is the sentence that hides the architecture.
- 3Demo Is Not DeploymentGoogle lost $100 billion of market cap in one afternoon because one demo slide said the wrong thing.
- 4The Nine Layers Where Agents Break"The AI failed" is not a diagnosis. It is a way to learn nothing.
- 5The Reliability EquationSame model, change containment, expected harm per claim drops 20x. Not the model.
Trust & Authority
What makes the AI coworker trustworthy?
- 6The Model Card Won't Save YouOpenAI retired its own headline benchmark. Of the tasks it audited, 59.4% had flawed tests.
- 7The Model Cannot Mark Its Own WorkAsking the model to grade itself is a faster way to find what it already believes, not what is true.
- 8Helpful, Harmless, and WrongA model can be aligned, pass every release gate, and still leave nobody accountable.
- 9Tools Give Models HandsOnce a model has tools, a wrong answer stops being words and becomes an action.
- 10The Supply Chain You Cannot SeeInstalling an agent component is not a trust ceremony. It is an authority transfer.
Runtime Control
What happens while it is running?
- 11The Sentence That Owns the AgentThe sentence that hijacks an agent is a polite line in a document you trusted.
- 12The Cheapest Token Is the One You Never GenerateMulti-agent work can burn ~15x a single chat. Cost is an authority design.
- 13What an Agent Remembers, and Cannot ForgetA customer asks to be forgotten. Has the agent forgotten them?
- 14Agents Don't Know When To StopMost production AI failures are confident handoffs at the wrong moment.
Traceability & Accountability
What does the run leave behind?
- 15Show Me the RunThe vendor says the agent completed the task. You say: show me the run.
- 16The Dashboard Is Green. The Meaning Is Wrong.Every check went green. The citations were perfectly formatted. They cited cases that do not exist.
- 17The Three Witnesses to a RunA model says it followed policy. The action log says retrieval failed. Three witnesses, one run.
- 18When an Agent Acts, Who Acted?An agent bought the wrong shoes. The charge cleared. Who acted?
- 19Compliance Is Not a PDFA regulator asks what your agent did between two and three. The vendor sends a thumbs-up.
Operating Model
Who decides what it does next?
- 20You Cannot Benchmark a CoworkerTwo green slides say the eval passed. They measured the wrong thing.
- 21The Autonomy LadderA weaker model with payment tools is more autonomous than a frontier model with none.
- 22Stop Delegating. Start Architecting.You are not delegating to a model. You are architecting a team.
Not sure where to start?
Ask Peter's profile AI to recommend a reading path for your role, or to summarise the operating model the series builds toward.