Four building blocks
World
A simulated system with services, data, and starting state.
Task
A job for your agent, with scorers that check success.
Agent
The AI agent you want to test.
Evaluation
A run that tests agents against tasks. Each attempt is a trial. In the dashboard, evaluations are called Backtests.
How agents are graded
Scorers check the agent’s actions, changed records, or final answer using code or an LLM judge. Only attached scorers affect the score, so add a check for each requirement you care about. Compare scores across agent versions to see what improved or regressed. Learn about tasks and scorers.Build, launch, evaluate
- Build: describe your system and review the world, tasks, and scorers.
- Launch: save a version and start its service twins—the simulated services your agent can call.
- Evaluate: run your agent, then inspect its scores and Timeline.
Build a world
Create a world and test your agent.