AI Agent Evaluation
Verify that your agent actually did the job
An agent that produces a confident summary of work it did not do looks identical to one that succeeded. People can tell the difference.
What workers do
- Decide whether the requested task was completed
- Identify the step where the agent went wrong
- Check whether the final output matches the request
- Flag actions that were unnecessary or unsafe
What you upload
- user_request
- actions_taken
- final_output
Any format: CSV, TSV, JSON, JSONL, XLSX or TXT. Columns are detected automatically.