Agent Reviewer Protocol
Task is done — now look back. What went well, what went wrong, what should be different next time? Goal: never repeat the same mistake and continuously improve skills and processes.
Core principle: Retrospectives are painful but necessary. A good agent evaluates itself.
6 Review Dimensions
1. Goal Alignment
Did the result match the original intent?
- Was the user's actual request met?
- Did scope creep occur?
- Over-delivery or under-delivery?
2. Efficiency
Did the task take longer than necessary?
- Unnecessary tool calls?
- Repeated operations?
- Sequential steps that could have been parallel?
- Token/resource waste?
3. Decision Quality
Were decisions well-reasoned?
- Were assumptions verified?
- Were alternatives considered?
- Did early decisions cause later problems?
4. Error Handling
How were errors addressed?
- Detected quickly?
- Right strategy applied?
- Same error repeated?
5. Communication
How was user interaction quality?
- Unnecessary confirmations requested?
- Critical information missing at key points?
- Too many or too few questions?
6. Reusability
Can lessons from this task transfer to the next?
- General patterns discovered?
- Which skills were missing or insufficient?
- Which decisions should become standard?
Finding Severity
| Severity | Meaning | Action |
|---|---|---|
| CRITICAL | Endangered the task or significantly reduced quality | Must fix |
| MODERATE | Created inefficiency but didn't break the result | Improve |
| POSITIVE | Something that went better than expected | Repeat, standardize |
Output Format
AGENT REVIEWER — Task Retrospective
Task : [task name]
Score : X/10
Findings : N critical | N moderate | N positive
## Dimension Scores
| Dimension | Score | Summary |
|-----------|-------|---------|
| Goal Alignment | X/10 | ... |
| Efficiency | X/10 | ... |
| Decision Quality | X/10 | ... |
| Error Handling | X/10 | ... |
| Communication | X/10 | ... |
| Reusability | X/10 | ... |
| **Overall** | **X/10** | |
## Critical Findings
[If any — what happened, why critical, how to prevent]
## Improvement Areas
[Inefficiencies, missed opportunities]
## What Went Well
[Decisions and approaches worth repeating]
## Action Items
### For Next Task
1. [Concrete change — what to do]
2. [Concrete change]
### Skill / Process Improvement
1. [Which skill should be updated / added]
2. [Which pattern should be standardized]
## Lessons Learned
[Items a future agent instance should know — candidates for memory-ledger]
Inefficiency Patterns — Auto-Detect
Scan the task history for these patterns:
| Pattern | Symptom | Fix |
|---|---|---|
| Repeated tool call | Same file/API read 2+ times | Cache it |
| Unnecessary confirmation | Low-risk step triggered approval | Adjust checkpoint-guardian threshold |
| Late assumption discovery | "Actually it should be..." after error | Trigger assumption-checker earlier |
| Sequential parallel steps | Independent steps ran sequentially | Use parallel-planner |
| Blind retry | Logic error treated as transient | Fix error-recovery categorization |
| Context loss | Previous step info forgotten | Memory-ledger not updated |
| Over-decomposition | 2-step task split into 8 | Adjust task-decomposer granularity |
Skill Performance Evaluation
Evaluate skills used during the task:
## Skills Used
| Skill | Used? | Effective? | Notes |
|-------|-------|------------|-------|
| task-decomposer | Yes/No | Good/Fair/Poor | ... |
| checkpoint-guardian | Yes/No | Good/Fair/Poor | ... |
| assumption-checker | Yes/No | Good/Fair/Poor | ... |
| tool-selector | Yes/No | Good/Fair/Poor | ... |
| parallel-planner | Yes/No | Good/Fair/Poor | ... |
| error-recovery | Yes/No | Good/Fair/Poor | ... |
| memory-ledger | Yes/No | Good/Fair/Poor | ... |
| output-critic | Yes/No | Good/Fair/Poor | ... |
Missing / untriggered skills and why?
When to Skip
- Task was single-step or under 5 minutes
- Prototype / experimental task
- User said "no retrospective needed"
Guardrails
- Be honest, not kind — the value is in finding problems, not hiding them.
- Concrete suggestions only — "do better" is useless; "cache file reads to avoid 3 redundant calls" is actionable.
- Cross-skill: this is the ecosystem's feedback loop — findings here should update other skills and processes.