Awesome Infra for AI › AI Safety & Guardrails

decionis/agent-safe-pipeline

⭐ 589 TypeScript added to this list on 2026-08-24 repository created 2026-08-13

Agent-Safe Pipeline implements the principle that an AI agent may propose an action but must not decide whether that action is authorized. The flow has four stages. Intent capture takes the agent proposal together with trusted context and freezes it into an immutable record, so what is evaluated is exactly what will run. That intent is submitted to an external policy service, Decionis, which returns one of three verdicts: allow, escalate or block. An escalation goes to a presence component that obtains verified human approval and returns the decision for re-evaluation rather than letting the approval bypass policy. Finally a SafeExecutor consumes a single-use grant bound to that specific intent and dispatches it through a sealed action registry that maps action names to trusted handlers and validates parameters. The executor never accepts an arbitrary callback from the agent, and the agent never holds the downstream privileged credentials or chooses which handler runs, which is what keeps the boundary meaningful. The repository is a runnable reference implementation and a set of packages rather than a hosted service, and it states plainly that its safety claims hold only while the documented trust boundary is preserved and that it does not replace provider-side identity, least privilege, network isolation or incident response. Packages cover the pipeline primitives, and examples walk from the smallest block flow upward. Supply chain hygiene is visible in the repository through continuous integration, code scanning, secret scanning and an OpenSSF Scorecard. It is aimed at teams giving agents the ability to call real APIs and needing a defensible record of who authorized each call.

https://github.com/decionis/agent-safe-pipeline

agent-authorizationpolicy-enforcementguardrailshuman-in-the-looptypescript