FULL STORY
SafeAgents: Hugging Face's Open Guardrail for Coding Agents
Hugging Face open-sourced SafeAgents, a local proxy that intercepts dangerous tool calls from coding agents, and quickly added pluggable decision-model checks and a log-only observation mode.
2026-10-29 ~ 2026-10-29 · 2 episodes · 5 posts
Episode 1 · Hugging Face Open-Sources SafeAgents to Intercept Risky Agent Tool Calls (2026-10-29, 3 posts)
Hugging Face open-sourced SafeAgents, a lightweight Python proxy that intercepts every tool call from coding agents before execution, using an ATIF trace format and around 110 deterministic detectors to block dangerous actions and reward hacking.
- safeagents: a lightweight proxy that blocks dangerous agent tool calls and reward hacking — yb2698 · 2026-10-29
- Inside safeagents: ATIF trace format plus ~110 deterministic detectors for tool calls — yb2698 · 2026-10-29
- SafeAgents: open-source local proxy blocks unsafe tool calls from coding agents — HauntingMoment · 2026-10-29
Episode 2 · safeagents adds pluggable decision models and log-only mode (2026-10-29, 2 posts)
safeagents now supports plugging in any decision model API for stronger checks on suspicious tool calls, and offers a --log-only mode that only logs without blocking, paired with a real-time terminal dashboard.
- safeagents adds --log-only mode: observe agent calls without blocking, with live dashboard — yb2698 · 2026-10-29
- safeagents plugs decision models into the loop, auditing only suspicious agent calls — yb2698 · 2026-10-29