Agent Evolve — Auto-Learning AGENTS.md Maintenance for Coding Agents
- URL: https://mp.weixin.qq.com/s/fmEZFYAFrsh-ZQaFSifMMA
- Date Saved: 2026-08-09
- Source: WeChat
- Tags: ai-engineering, ai-tools
Summary
Introduces “Agent Evolve” — a Harness component from Agent Plan (火山方舟) that automatically learns from past agent sessions and generates improvements to AGENTS.md/TOOLS.md instruction files. Solves the pain of manually maintaining agent config files as projects evolve.
How it works (3-step loop)
- Auto-review sessions — identifies repeated corrections, error patterns, reusable experience from recent task records
- Generate improvement suggestions — converts experience into specific diffs for AGENTS.md, TOOLS.md
- Apply after confirmation — human reviews diff, approves, then auto-writes back
Real-world benchmark
Same web scraping task, before vs after Evolve:
| Metric | Before | After |
|---|---|---|
| Time | 12.7 min | 46.8 s |
| LLM calls | 30 | 6 |
| Tool calls | 28 | 5 |
| Token consumption | 4.29M | 321K |
(Note: environment setup was already done on first run; second run benefited from reduced wrong tool calls)
5 Rules Evolve Generated
Written to AGENTS.md:
- Goal drift detection — every 3 steps, check alignment with original user goal
- Entity normalization — unify names, preserve number precision, deduplicate
- Structured data parsing verification — compare extracted fields with raw data
- Planning depth expansion — analyze preconditions, output usage, fallback plans before multi-step tasks
Written to TOOLS.md: 5. Tool parameter schema pre-validation — check types/formats before calling (dates as YYYY-MM-DD, no units in numbers, enum values only)
Setup
Supports: Claude Code, OpenClaw, TRAE
- Get Agent Plan API Key from 火山方舟 (Volcengine)
- Install Evolve Skill:
curl -fsSL "https://ark-self-evolve.tos-cn-beijing.volces.com/evolve_skill/latest/install.sh" | bash - Tell agent: “Set me up for evolve”
Usage:
- “Learn from my recent sessions” → extracts patterns
- “Show me the diff” → preview changes
- “Apply this proposal” → write back