Monitors and control protocols that watch agents and intervene in their execution.
A monitor observes a step of an agent's execution and reports a suspicion score; it never acts. A protocol decides what happens at that step, optionally after consulting monitors, and a protocol with no monitors is a rule. Both run at one of four stages, the points in the agent loop before or after a model generate and before or after a tool call. The same monitor runs inside an Inspect eval, in a network proxy in front of a model API, and offline over recorded transcripts for validation and calibration.
pip install git+https://github.com/meridianlabs-ai/inspect_sentinelThis also installs the development version of inspect_ai from its feature/sentinel branch, which Sentinel tracks until that branch merges.
The package is under active design. The design documents are in design/; start with sentinel-overview.md.
Runnable examples are in examples/.
Documentation: https://meridianlabs-ai.github.io/inspect_sentinel.