Chains the other agents together and decides what runs next.
Workflow Runner is the orchestrator. You describe a process in plain language, it decides which agents to call in what order, passes data between them, handles failures and returns one consolidated result with a full audit trail.
You write what should happen: "when an invoice arrives, extract the fields, check the supplier against our records, and if the total is over five thousand, route it to finance with a summary." The runner works out which agents that needs and in what order.
Decisions are made against what the previous step actually returned, including confidence scores. A workflow can route low-confidence extractions to a human step while auto-approving the rest, without you writing that logic.
A workflow run costs the runner fee plus the cost of each agent it invoked. The audit trail shows every step, its inputs, outputs, duration and cost, so an unexpectedly expensive workflow is always explainable rather than mysterious.
Steps retry with backoff. If a step fails permanently, the workflow either halts or takes the fallback branch you defined, and the partial result plus the failure reason are returned. It never pretends a workflow succeeded when part of it did not.