Agent Evaluation Lab

See how a measured agent run is evaluated on outcome and trajectory separately — task success is not the same as a good tool path.

GuidedNo API keys in LabAI Agents

Example evaluation

Select a measured Cookbook case, then Run to load its recorded trace and computed evaluation.

Task: Check the payments service. If it is degraded, inspect the relevant documentation and summarize what the user should know.

Example evaluationTask successCorrect tools, arguments, observations, and answer — outcome and trajectory both pass.Partial successCorrect outcome after an extra tool call that violates this case’s efficiency constraint.Tool error recoveryA failed tool call is recovered, then the task completes. The failure stays in the trace.Goal missThe expected tool path is followed, but the final answer is incorrect.

Choose an example evaluation and run the recorded Cookbook case.

Related Labs

Tool calling and the agent loop produce traces; this Lab evaluates those traces. Planning is the next runtime lesson.

02 Agent Loop03 Agent Evaluation04 Planning

Continue Learning

Guides that explain outcome, trajectory, and evaluation.

Implementation

Run the measured Cookbook example locally.

Agent Evaluation Cookbook

Related Tools

Tools commonly used with agent evaluation stacks.

Related Rankings

See how this stack compares in DataAIHub rankings.

Best AI Agent Frameworks

Lab ID agent-evaluation · guided / precomputed