Core concepts
Jobs and runs
Define repeatable work and understand each execution.
A job combines an agent with a specific task prompt. It may also have one trigger. The job is the reusable definition; a job run is one execution of that definition.
Each run records its status, trigger source, messages, model calls, tool calls, human-input requests, output, errors, and timestamps. A run may be pending, running, awaiting input, completed, failed, or cancelled.
Use one job per stable outcome. Put durable role and safety instructions on the agent; put sources, destination, cadence-specific context, and expected output on the job.