Step 1 / 5 · 15 min
Define when the agent starts and what completion means.Contract the job
- Choose a repeated, bounded job.
- Specify eligible input and an inspectable output.
Build · Core skill lab
Workflow Contract + Dry Run
Design a tool-using loop with visible checkpoints, bounded authority, and a stop condition before automating it.
- Decision
- Can this workflow run repeatedly without relying on the agent to invent policy?
- Timebox
- 90 minutes + one manual dry run
- Output
- Agent workflow contract
Bring to the bench
- A repeated operational job
- Known tools and data sources
- A human owner for escalation
Your work saves in this browser.
Field tools
Use the instrument, not a blank page.
Copy these into your interview, agent, review, or working document. They are specific to this repetition.
template
Workflow contract
This is the minimum policy surface for a repeatable agent loop.
TRIGGER / ELIGIBLE INPUT OBJECTIVE / OUTPUT SCHEMA ALLOWED TOOLS + READS ALLOWED WRITES FORBIDDEN ACTIONS CHECKPOINTS + ARTIFACTS QUALITY GATE STOP CONDITIONS ESCALATION OWNER + PARTIAL HANDOFF AUDIT LOG FIELDS
checklist
Dry-run probes
Pause at each checkpoint and interrogate the contract.
Did the agent need information it was not authorised to fetch? Did a tool result require domain judgment? Could an incorrect assumption trigger a write? Is progress visible before completion? Can another human reproduce the decision? Does stopping preserve useful partial work?
Calibrate judgment
Compare the evidence, not the polish.
A workflow that knows when to stop
An incident-triage agent reads issues, logs, and deploy data.
The agent can read redacted evidence and draft a brief, but cannot deploy or contact customers. It checkpoints facts, hypotheses, and verification. After 20 minutes or two failed queries it hands the partial brief to the incident lead.
Why it works: Authority, observability, quality, and partial handoff are designed rather than delegated to improvisation.
Autonomy as instruction
Monitor incidents, investigate problems, use any tools you need, fix issues where possible, and notify the team when done.
Why it fails: Trigger, permissions, consequence boundaries, checkpoints, quality, and escalation are all undefined.
Review → revise → repeat
The artifact is the beginning of the rep.
Check only standards your current artifact actually meets. Then record one consequential revision before exporting it.