Skip to content
Donnaby The Architect LabBook a free call
Book a free call

How we work

A fairly standard engineering workflow. Just with AI agents.

Define the problem. Build in isolation. Prove it. Have someone else check it. Ship it. Watch it. Every competent engineering team has worked this way for twenty years. The difference now is who does the steps.

Humans own two gates: the problem and the release. Agents do the rest and are checked by other agents.

The loop

Nine steps · two human gates · two feedback loops

The Architect Lab engineering workflow with AI agentsNine steps. Idea, Specify, Isolate, Build, Prove, Review, Approve, Ship, Monitor. A failed review returns work to Build. Monitoring returns new work to Specify. A human defines the problem and approves the release. Not approved means hold.Spec and buildCheck and release01 · HumanIdeaHuman definesthe problem02 · Human + agentSpecifyRequirements andacceptance criteria03 · AgentIsolateOwn workspaceand branch04 · Engineering agentBuildAgent implementsthe spec05 · Engineering agentProveTests, screenshots,logsEvidence06 · Review agentReviewIndependent check.Passes?07 · HumanApproveHuman releaseapproval?08 · AgentShipMerge and deploy09 · Agent + humanMonitorErrors, feedback,improvementsYesApprovedNot approvedHoldDo not releaseNo · fix and prove againNew work
  • Human gate
  • Decision
  • Forward
  • Feedback loop
Two feedback loops. A failed review returns the work to the engineering agent. Monitoring returns new work to Specify.

Who does what

Three roles · no overlap

01 / Human

Decides

Defines the problem. Signs off the spec. Approves or holds the release. Never writes the code and never reviews a diff line by line.

  • Owns the acceptance criteria
  • Owns the release decision
  • Reads evidence, not code

02 / Engineering agent

Builds and proves

Works in an isolated workspace and branch. Implements the spec. Produces the evidence that it works: tests, screenshots, logs.

  • One task per workspace
  • No access to production
  • Burden of proof is on the builder

03 / Review agent

Checks

Has not seen the build. Gets the spec, the diff and the evidence. Passes or fails the work against the acceptance criteria. Nothing to defend.

  • Separate context, separate session
  • Fails anything without evidence
  • Returns failures to the builder

The nine steps

Same loop for software, content and operations

  1. 01

    Idea

    A human defines the problem. What is broken, who it affects, what good looks like. No solution yet.

    Human gate

  2. 02

    Specify

    Requirements and acceptance criteria written before any code. The spec is the contract the agent builds against and the reviewer checks against.

    Human · Agent

  3. 03

    Isolate

    Every piece of work gets its own workspace and branch. Nothing touches shared code or production until it has been reviewed.

    Agent

  4. 04

    Build

    The engineering agent implements the spec inside the isolated workspace. Scope is the spec. Anything outside it is a new idea.

    Engineering agent

  5. 05

    Prove

    The same agent produces evidence. Passing tests, screenshots of the running feature, logs from the run. No evidence, no review.

    Engineering agent

  6. 06

    Review

    A separate review agent checks the work against the spec and the evidence. Fail sends it back to Build. Pass sends it to a human.

    Review agent

  7. 07

    Approve

    A human decides whether to release. Approved goes to Ship. Not approved goes to Hold. The human reads the evidence and the review, not the code.

    Human gate

  8. 08

    Ship

    Merge and deploy. Automated. The branch closes, the workspace is discarded.

    Agent

  9. 09

    Monitor

    Errors, feedback and improvement requests are tracked. Anything worth doing becomes new work and enters at Specify. Shipped is not finished.

    Agent · Human

What never changes

Six rules · every project

  1. 01A human writes the problem and the acceptance criteria.
  2. 02Every task runs in its own workspace and branch.
  3. 03The agent that built it has to prove it. Tests, screenshots, logs.
  4. 04A separate agent that has not seen the build does the review.
  5. 05A human approves every release.
  6. 06Shipped is not finished. Monitoring feeds the next spec.

Questions

Four · short answers

Why does a separate agent review the work?
An agent grading its own work passes its own work. The review agent has not seen the build. It gets the spec, the diff and the evidence, and it has nothing to defend.
What counts as evidence?
Passing tests, screenshots of the feature running, logs from the run. If the agent cannot show it, it did not happen.
Where do humans sit in the workflow?
Two gates. A human defines the problem at the start and approves the release at the end. Everything between is done by agents and checked by agents.
Does this only apply to software?
No. The same loop runs a content pipeline, a CRM rebuild or an operations automation. Specify, isolate, build, prove, review, ship, monitor. The tools change. The gates do not.