Runs and steps

A run is one execution of an agent definition caused by one event. A pipeline runs its agents in order; each agent sees the event and the output of the previous agent.

States

queued ──▶ running ──▶ succeeded
   │          │  ▲ └──▶ failed | cancelled | blocked_by_policy
   │          ▼  │
   │   awaiting_approval ──▶ failed | cancelled | blocked_by_policy
   └──▶ cancelled

succeeded, failed, cancelled and blocked_by_policy are final. A run waits in awaiting_approval when a tool that requires approval is called; approvers have approvals.timeoutSeconds (default 3600) to decide.

Steps

Each step has a kind and a status:

Kind	Meaning
model_call	A request to the model provider, with token usage and cost
tool_call	An MCP tool call that passed the gate
policy_decision	The audit gate’s decision for a requested call
approval	A human approval request and its outcome
control	A control agent action (pause, kill)
output	A result in one of the agent’s output formats
error	An error with a stable code

Statuses are ok, error, denied, pending, approved and rejected.

Configuring a run

Budgets, approvals, timeouts and the runner are declared in agents.md (see reference). Pipeline and agent budgets are merged per field and the stricter value wins.