Agent Operating Manual Template
The one-page manual for a production agent: mission, capabilities, boundaries, escalation, and ownership. Copy it below, fill the bracketed fields, delete what you don't need.
# Operating manual: [agent name]
Owner: [named human] · Review cadence: [monthly]
## Mission
[What this agent is for, in one sentence a stakeholder would recognize.]
## Capabilities
- [What it does, honestly scoped]
## Boundaries
- Never: [enumerated forbidden actions]
- Requires human approval: [enumerated gated actions]
- Permission profile: [tools, paths, network access]
## Escalation
- The agent escalates when: [budget exhaustion, gate hits, uncertainty]
- Escalations go to: [channel/person] with [evidence attached]
## Failure playbook
- If it produces wrong output: [who is paged, how to correct, where corrections are logged]
- Kill switch: [how to stop it, who may]
## Audit
- Actions logged to: [hook-written log location]
- Log reviewed: [cadence, by whom]
Field notes
| Field | Why it matters |
|---|---|
| Owner | A named human, not a team alias. Accountability that resolves to nobody resolves to nobody exactly when it matters. |
| Kill switch | Write down how to stop the agent before you need to know. This line has the best worst-case ROI in the document. |