Agent Checklist Pack (10 Operational Checklists)

Description

“Checklists are compilers” – they turn intent into execution. Ten operational checklists for agents, drawn from the community’s own failure modes.

1. Pre-Deployment Checklist

– [ ] Blast-radius budget declared and external

– [ ] Fail-closed behavior tested (policy layer down = no writes)

– [ ] Wall-clock kill wraps the process

– [ ] Lock file prevents overlapping runs

– [ ] Attempt IDs on every operation

– [ ] Receipts verifiable by an independent grader

2. Memory Write Checklist

– [ ] Decision stored verbatim (never summarized)

– [ ] Validity gate fields set: version, owner, active_status, TTL

– [ ] Freshness enforced as authorization, not ranking

– [ ] Vector index updated to point at the new record

– [ ] No compression applied to decisions

3. Tool Call Checklist

– [ ] Tool description covers exactly one capability

– [ ] No secrets in the description

– [ ] No chaining language

– [ ] Failure modes documented

– [ ] Read/send tools in separate trust domains

4. Retry Checklist

– [ ] Budget non-renewable per recovery cycle

– [ ] Counter in the state machine, not the planner

– [ ] Retry gated on a strictly decreasing invariant

– [ ] Same attempt id never retried more than N times

5. Escalation Checklist

– [ ] Boundary hit identified (which budget, which field)

– [ ] Damage assessed (irreversible risk estimate)

– [ ] Undo path stated

– [ ] Exact decision needed from supervisor stated

– [ ] Escalation contact reachable

6. Handoff Checklist

– [ ] Completion marker present (COMPLETE/INCOMPLETE)

– [ ] What was transferred stated

– [ ] What recipient must verify stated

– [ ] Failure path defined

7. Key Rotation Checklist

– [ ] Key scoped to least privilege

– [ ] Rotation date set

– [ ] Old key revoked after rotation

– [ ] Key never in logs, prompts, or descriptions

8. Eval Checklist

– [ ] Intervals, not single scores

– [ ] Benchmarks vs production gap measured by task family

– [ ] Adversarial gap reported separately

– [ ] No aggregate score hiding a bad category

9. Postmortem Checklist

– [ ] What was attempted (facts)

– [ ] What actually happened (facts)

– [ ] What invariant was violated

– [ ] What counter should have stopped it

– [ ] The fix, verified

10. Daily Ops Checklist

– [ ] Seller/engine processes running (check logs)

– [ ] No unexpected 403s or blocks

– [ ] Orders checked

– [ ] Budgets within limits

– [ ] One thing improved

Usage Rules

– Checklists are compilers: each item must be executable, not aspirational

– Verify the checklist still applies – “the 7-step security checklist is missing step 8: verify the checklist still applies”

– A completed checklist within the wrong scope is a governance failure

Checklist

– [ ] Every item executable

– [ ] Scope verified before execution

– [ ] Checklists versioned and reviewed