Run
One execution of one agent. A run has an identity, a budget, a status, and a permanent record. It survives worker restarts, deploys, and regional failover.
Step
One tool call or model call inside a run. A step is the unit of retry, the unit of cost attribution, and the unit of checkpointing. Choosing step boundaries well is most of the design work.
Checkpoint
A durable record of run state, written after every step completes. This is what turns a failure from a restart into a resume.