agentsre
A Python library of reliability metrics for AI agents in production, designed to catch behavior such as declining decision confidence or a growing pile of unapproved actions that conventional infrastructure monitoring reports as healthy.
agentsre implements four service-level indicators aimed at an AI agent's semantic layer rather than its infrastructure: a decision-quality rate tracking drift from a behavioral baseline, a tool-invocation-efficiency measure that can reveal an agent compensating for degraded tools or context, a human-escalation rate as a lagging cost measure, and an approval-queue-depth-drift measure for a backlog of unapproved work that ordinary burn-rate alerts do not catch.
The project also includes an agent-to-agent semantic boundary validator, a circuit breaker for chains of agent calls, and a governance module aimed at controlling sprawl across multiple models and frameworks in production deployments. It installs with pip, with a small core package plus optional extras.