We built Geneclaw, an open-source AI agent framework with a built-in self-evolution engine. The core idea: instead of only executing tasks for users, the agent continuously observes its own runtime failures, diagnoses root causes, generates constrained code patches, and (optionally) applies them — all behind a 5-layer safety gatekeeper.
Key design decisions:
- Dry-run by default. Nothing is applied without explicit human approval.
- 5-layer Gatekeeper: Path allowlist/denylist, diff size limits, secret scanning, and code pattern detection (blocks eval/exec/os.system).
- Git safety: Every patch creates a dedicated branch, runs `git apply --check` first, executes tests after patching, and automatically rolls back on failure.
- Full audit trail: Every event in the evolution lifecycle is logged as append-only JSONL with secret redaction. A Streamlit dashboard visualises the data.
- LLM-optional: Works in heuristic-only mode without any LLM provider configured.
geneclawai•1h ago
We built Geneclaw, an open-source AI agent framework with a built-in self-evolution engine. The core idea: instead of only executing tasks for users, the agent continuously observes its own runtime failures, diagnoses root causes, generates constrained code patches, and (optionally) applies them — all behind a 5-layer safety gatekeeper.
Key design decisions:
- Dry-run by default. Nothing is applied without explicit human approval. - 5-layer Gatekeeper: Path allowlist/denylist, diff size limits, secret scanning, and code pattern detection (blocks eval/exec/os.system). - Git safety: Every patch creates a dedicated branch, runs `git apply --check` first, executes tests after patching, and automatically rolls back on failure. - Full audit trail: Every event in the evolution lifecycle is logged as append-only JSONL with secret redaction. A Streamlit dashboard visualises the data. - LLM-optional: Works in heuristic-only mode without any LLM provider configured.
Geneclaw is built on top of nanobot (https://github.com/HKUDS/nanobot), a lightweight personal AI agent framework from HKU.
The evolution pipeline: Observe → Diagnose → Propose → Gate → Apply
Architecture and quick start: https://geneclaw.ai Protocol spec: https://github.com/Clawland-AI/Geneclaw/blob/master/docs/spe...
We'd love feedback on the safety model — is 5 layers enough? Too much? What would you want before letting an AI agent modify its own code?