What is a Tessary classifier?

A classifier scores one trace for one property: a number and a confidence, not a verdict on whether the agent did its job. Each one is a narrow, always-on check with negligible per-trace cost, which is what makes it affordable to run against all of your traffic instead of a sample. The drift checks judge a trace against a baseline learned from that call site’s own recent history rather than a fixed number, so what counts as unusual is specific to your agent.

Built-ins cover secret_leak, malformed_output, frustration, groundedness, behavior_drift, duration_drift, cost_drift, and tool_error. Custom classifiers, tuned to your own agent’s failure modes, run alongside them.

Each classifier reads one signal, not the whole turn. It can tell you a user sounds frustrated; it can’t tell you whether your agent should have declined the request.