Agent Retrospective
Vetted and production-tested agent skill.
Agent Retrospective
Self-improvement loop for AI agents: trace every completed task, judge with a different model, diagnose patterns, and apply concrete fixes.
How It Works
After every task completion, the system:
1. **Captures a trace** — metrics, patterns, outcome (disk only, never resident)
2. **Judges with a different model** — isolated sub-agent reviews the trace
3. **Writes improvement signals** — concrete, fixable issues only
4. **Hill-climbs weekly** — batches signals, validates, proposes fixes
Architecture
TASK COMPLETE → trace capture → judge review → signal queue → weekly hill-climb → human approval → apply fixEvery step stays on disk until a human approves the change.
Key Principles
- **Judge ≠ maker** — never let the same model grade its own work
- **Only concrete fixes** — no "be more careful" signals. Specific problem, evidence, proposed fix, target file
- **Memory discipline** — traces and signals on disk. Zero resident token bloat
- **Human gate** — skills, prompts, config → human approval required
What It Improves
| Pattern | Example Signal |
|---|---|
| Tool misuse | Agent re-fetches truncated content 3x before raising limit |
| Context loss | Agent drifts from spec in long sessions |
| Model choice | Expensive model used for simple sub-agent work |
| Skill gaps | Missing error handling pattern in a skill |
| Cost waste | Redundant tool calls eating token budget |
Results
Over time, agents genuinely get better — fewer repeated mistakes, smarter tool usage, compound optimization across skills and memory.
Files
scripts/task-trace.sh — Captures completion metrics
scripts/judge-task.sh — Stages isolated judge review
scripts/hill-climb.sh — Weekly batch processor
memory/improvements/ — Signal storage (disk only)
skills/agent-retrospective/ — Skill definition + judge promptGet the full skills pack
Enter your email to download all 15 AgentVet skills as a pack. No spam, just occasional updates on new skills.