agent-evolvelisted

# /agent-evolve — Agent Self-Improvement Loop > The agent that runs the agents needs a meta-loop: am I getting better, or am I just running? ## What it does Periodically — weekly, monthly, after a notable failure pattern — run this loop to identify and apply improvements to the agent's prompts, configuration, and operational guardrails. Six phases: 1. **Health verify** — is the agent actually working? `/agent-status` first. 2. **Pipeline analysis** — what does failure data over the last window reveal? 3. **Pattern research** — what's the current state of the art for autonomous coding agents? 4. **Proposal** — what specific changes (to prompts, config, guardrails) would move the metrics? 5. **Apply** — make the changes (with operator approval). 6. **Measure** — track whether the changes moved the metrics. ## Prerequisites | Var | Purpose | |---|---| | `AGENT_API_URL` | Base URL of the agent's HTTP API | | `AGENT_API_TOKEN` | Bearer token | | `AGENT_REPO_PATH` | Local path to the agent's source code (so changes can be applied via git) | ## Phase 1: Health verify Run `/agent-status` first. If the agent isn't healthy, **stop here**. Don't try to evolve a broken agent — fix it first. ## Phase 2: Pipeline analysis Pull data for the analysis window (default: last 7 days): ```bash # All tasks in window curl -s -H "Authorization: Bearer $AGENT_API_TOKEN" \ "$AGENT_API_URL/api/tasks?since=7d" ``` Compute: | Metric | Why it matters | |---|---| | Fail rate | Trend up / do