agent-safe-pipeline

Why we included this project

Agents usually get to judge whether their own actions are safe; this project is built on the opposite idea. It is a reference architecture and library that separates proposing from authorizing: the agent's intended action is captured as an immutable, hashed intent, an independent Decionis policy service returns an ALLOW, ESCALATE, or BLOCK verdict, and a SafeExecutor runs only the actions that match a single-use, intent-bound grant. Human approval goes through a separate verified path that the policy engine re-checks, so autonomy does not have to mean handing the model the keys. The repo ships runnable examples covering refunds, deploys, and procurement, plus an MCP tool gate and threat-model docs. It is a library, not a hosted service: production use still calls for provider-side identity, least privilege, network isolation, and incident response.