Own Your AI Workflow: Contracts, Tests and Human Decisions

Define one observable behavior: a saved prompt survives reload and appears with the correct format. Keep the task bounded. No model name or current-year framing makes acceptance criteria unnecessary.

Provide the relevant files, expected input/output, allowed changes and stop conditions. Separate source facts from suggestions. Do not ask an assistant to infer production credentials or permission to publish.

Inspect the diff and run the behavior check. Test unavailable records and failed requests as well as the happy path. Preserve existing bookmark keys. Log unresolved issues instead of calling a passing build complete evidence of usability.

The owner sets scope. The assistant proposes and implements local changes. A reviewer checks behavior and evidence. The owner authorizes deployment separately. Adding more assistants does not remove accountability.

Record task completion, escaped defects and maintenance effort before claiming improvement. A functional success is not proof of tenfold productivity, traffic growth or revenue.

• Agent workflow design (foundational, December 2024)

Use a fictional weekly update: two approved facts, one unresolved decision and one deadline. The task is to draft a reviewable update; the assistant must not send it. A single assistant and a checklist may be enough. Add a second role only when it performs an independently useful check.

Approved facts: prototype delivered; accessibility review not completed.
Unresolved: release owner.
Deadline: review draft by Thursday.
Deliverable: five-sentence internal update plus a list of unknowns.
Forbidden: announce launch, assign the owner, claim accessibility passed.

The reference check is simple: delivery can be stated, accessibility remains pending, the owner remains unknown and Thursday is the review deadline. A confident launch announcement fails the task.

A drafting role reads only the approved facts. A checking role receives the same facts and the draft, returning supported, unsupported and missing claims. The human owner resolves unsupported content and decides whether to send. Both roles sharing the same model is not independent external verification; check the original facts yourself.

Keep a stable task ID, save draft versions and record who approved the final update. On timeout, leave the task pending. On a duplicate request, reuse the same task rather than sending a second update. On conflicting feedback, return the conflict to the owner; do not let another agent silently choose a policy.

Compare this design with a single-assistant run using the same input. Count review corrections and total effort, not just generation time. Keep the simpler design if the extra role adds cost without catching useful errors. No model benchmark or productivity gain is claimed here.

Bound the work with an agent contract · Separate proposals from meeting decisions