Agent Guard uses CEL policies to block coding agents' risky dependency installs
HowDevelop · x · 2026-10-04
Coding agents can autonomously run npm install or pip install, turning bug fixes into supply-chain decisions with potentially hallucinated package names. Jozu AI's Agent Guard addresses this: apply a ToolPolicy to the agent's shell tool (e.g. Bash), evaluate command arguments against CEL rules, and set action to Enforce so failed assertions block the tool call before execution. The author warns that command matching needs care — rules targeting pip install may miss python -m pip, uv pip, or other wrappers.
More from coding & agent
- Dev Keeps Asking Codex 'Wait, How Did You Do That?' as Agent Surprises Him — yungcontent · 2026-11-20
- Dev deletes all ADRs: 49% superseded in days, codebase is now your source of truth — sloppenheimer · 2026-11-20
- Anthropic opens Claude Forward Deployed Engineer learning path, ~9.5 hours self-paced — leslysandra · 2026-11-20
- Google open-sources Sashiko, a Gemini agent that cut Linux post-merge fix rates 35% — rseroter · 2026-11-20
- boxd adds 8 more integrations including Dify, n8n, LangChain, and Neon — JensHonack · 2026-11-20
- OpenAI updates prompting guidance for live-1.1: languages and regional accents — pbbakkum · 2026-11-20