Agent Reliability Review Template
A self-audit of an agent system against the Agentic Maturity Model: failure inventory, loop and memory audit, top fixes. Copy it below, fill the bracketed fields, delete what you don't need.
# Agent reliability review: [system]
Date: [date] · Reviewer: [name]
## Maturity level (be honest)
- What happens when the agent is wrong? [nothing catches it / a loop catches it / caught and remembered / caught, remembered, auditable]
- Claimed level: L[0-5] · Evidence: [one sentence]
## Failure inventory (last 30 days)
| Failure | Frequency | Layer (identity/skill/loop/memory/eval/governance) | Cost |
|---------|-----------|-----------------------------------------------------|------|
| [what went wrong] | [n times] | [layer] | [hours/incident] |
## Loop audit
- Does anything refuse to accept bad work? [what, specifically]
- Verification is: [evidence-based / assertion-based / absent]
- Budgets and abort paths exist: [yes/no, where]
## Memory audit
- Facts dated: [yes/no] · Re-verified on recall: [yes/no]
- Last consolidation pass: [date / never]
- Write rules exclude secrets: [yes/no]
## Top 5 fixes, ordered by failure-cost
1. [fix] - addresses [failure], builds [layer]
2-5. [...]
## Next review: [date]
Field notes
| Field | Why it matters |
|---|---|
| Maturity level | The what-happens-when-wrong question locates you in one pass and resists flattery. Most teams claiming L3 discover assertion-based verification here. |
| Failure inventory | Layer attribution turns complaints into a build order: the layer with the most expensive failures is your next investment. |