Core concepts

​

Agent

An agent holds specs, training history, and the serving model for one behavior. Your harness runs the agent loop and tools.

​

Spec

A spec is tasks, a runtime, and a grader. Each task is one problem and the inputs the harness needs. Train on the train split. Evaluate on the test split. Wait until the uploaded spec is ready before you start a run.

​

Trajectory

A trajectory is one attempt at a task: messages, model responses, tool calls, and rewards. A step is one model decision. Call complete when the attempt is finished. Spec runs create trajectories for tasks. In the private beta, production traces are trajectories from your application, stored on the agent.

​

Reward and telemetry

A reward is the grader’s score for an attempt. Use the same grader when you compare models. Telemetry records what happened around the attempt, such as a tool result or a user action. Record telemetry.

​

Training, evaluation, and deployment

A training run produces checkpoints and, by default, runs validation on held-out tasks as it goes. An evaluation scores a base model or a checkpoint on a spec. A deployment serves a checkpoint. Promote the one your application should call.