agent-seed

Build the strongest model-agnostic agent harness through self-improvement.

A self-improving autonomous daemon that runs a while-true loop: read state, call LLM, execute actions, validate changes, commit, health-check, repeat.

Architecture

daemon.py          — merged main loop (~200 lines)
agent_session.py   — lightweight API wrapper, model routing from .model-config.json
safety.py          — 5 safety layers (blocked commands, disk quota, JSON validation, health check)
state.py           — reads GOAL.md + git state + eval results
git_workflow.py    — commit, push, rollback, gc

Configuration

Edit .model-config.json to set up model providers and routing:

"providers": {
    "deepseek": { "type": "api", "base_url": "https://api.deepseek.com/v1", "api_key_env": "DEEPSEEK_API_KEY" },
    "local":    { "type": "local", "base_url": "http://localhost:11434/v1" }
}
"routes": {
    "create":  { "model": "deepseek-chat", "provider": "deepseek", "access": "read_write" },
    "explore": { "model": "qwen3.6:27b", "provider": "local", "access": "read_only" }
}

Set API keys via environment variables (e.g. DEEPSEEK_API_KEY).

Running

python3 daemon.py

Environment variables:

• AGENT_SEED_SLEEP — base sleep between cycles (default: 60s)
• AGENT_SEED_SLEEP_FAIL — sleep after failure (default: 300s)
• AGENT_SEED_MAX_FAILURES — consecutive failures before escalation (default: 3)
• AGENT_SEED_EXPLORE_INTERVAL — every Nth cycle is EXPLORE (default: 4)
• AGENT_SEED_PUSH — auto-push to origin main (default: true)
• DEEPSEEK_API_KEY — API key for DeepSeek (if using deepseek provider)

Safety

5 safety layers, enforced at the Python level (never trust the model):

1. Blocked commands — rm -rf, git reset --hard, dangerous patterns
2. Blocked paths — daemon.py, safety.py, AGENTS.md, GOAL.md etc.
3. Disk quota + log rotation — checks free space, compresses old logs
4. JSON schema validation — validates config file changes, auto-reverts
5. Health check + rollback — post-commit checks, git rollback on failure

Scripts

• scripts/eval — self-evaluation framework
• scripts/improve — state aggregation for improvement decisions
• scripts/route — model route resolver
• scripts/go — iteration protocol reference
• scripts/commit — safe commit wrapper

Tests

bash tests/smoke.sh