Agentic Evaluation
Evaluate AI agents in production: task completion metrics, trajectory analysis, and automated agent quality benchmarks.
Last updated
After this section you can
- Evaluate an agent on both axes — outcome (did it work) and trajectory (how it got there)
- Build an eval harness for multi-step agent runs
- Carry offline evaluation through to production monitoring
Grading Agents: Agentic Evaluation
Evaluating an agent is not like evaluating a model. There’s no single output to score — there’s a non-deterministic, multi-step trajectory that can reach the right answer the wrong way, or the wrong answer competently. This is how you grade the difference.
A one-shot model gives you one output; you score it against a reference and move on. An agent takes a path — it plans, calls tools, reads results, re-plans, and stops when it decides it’s done. Two runs of the same task take different paths, and both “84% correct” can hide wildly different behavior. So agent evaluation splits in two: outcome (did it achieve the goal?) and trajectory (was the path sound — efficient, safe, non-looping?). You need both, because an agent that gets the right answer by luck after twelve wrong tool calls is a production incident waiting to happen.