Self Improving Agent

Always-on lightweight observer. After EVERY user message, silently check: (1) did the user correct me or express a preference? (2) did the user express frustration about repetition? (3) is this a rule I've seen before in a different project? If ANY are true, run `uv run improve.py observe "<rule>" --project <name> [--domain <domain>]` to record it. Do NOT load the full skill for observation — this description is sufficient. Load the full SKILL.md body only when: checking escalation thresholds, writing to AGENTS.md, or running /improve. Manages ~/.agent-improvement/rules.yaml via improve.py CLI.

ericmjl Updated 41 repo stars
Junpeng Lao: Really like opencode-autolearn, it's the worklog/self-improvement loop done properly.

File contents

ericmjl/opencode-autolearn commit 4f8f38c102

Frequently asked questions

npx get-skill add ericmjl/self-improving-agent