AgentGuard: an open-source deterministic permission layer for AI agent tool calls
Glittering-Coat-657 · reddit · 2026-09-02
A Reddit developer open-sourced AgentGuard, a tiny permission layer that sits between an agent and its tools (agent → AgentGuard → tool) and is now inviting developers to try to break it on GitHub.
- Checks whether a tool is allowed in the current agent state, whether arguments fall within policy, and whether the state is unknown/unsafe, and verifies the underlying function stays untouched when denied
- Example denials: researchagent calling refund or deletedatabase, refundagent issuing a $1,000 refund
- Deliberately minimal design: no dashboard, no cloud, no "AI judging AI" — just deterministic execution-time policy
- Author seeks developers on LangGraph/LangChain/MCP agent stacks to stress-test bypasses and gauge real-world installability and willingness to pay
More from coding & agent
- Submitting a PR Isn't a Flex Anymore in the AI Coding Era — pjausovec · 2026-09-02
- Open-source 7-layer production agentic AI system hits 937 GitHub stars — mdancho84 · 2026-09-02
- Jest creator Nakazawa: six engineering values that matter more as coding agents speed up — bibryam · 2026-09-02
- ODS turns your PC into a private AI server with Ollama, n8n and ComfyUI pre-wired — tom_doerr · 2026-09-02
- Reverse-engineering ChatGPT's memory: 4 layers, no vector DB, no RAG — HowDevelop · 2026-09-02
- Google's $300 credit broke Gemini nodes in ComfyUI, so one dev built a Vertex AI workaround — Cute-Appointment6874 · 2026-09-02