Workflows

A workflow is a declarative finite state machine that coordinates multiple steps — agent invocations, container runs, human approvals, parallel fan-outs, and nested subworkflows — into a single, replayable process. Where an agent does one thing well, a workflow strings many agents and systems together to deliver a higher-level outcome.

Workflows are defined in YAML, stored in the workflow registry, and executed by the orchestrator. Every step reads from and writes to a shared blackboard — a mutable JSON context that flows through the run — and transitions are governed by explicit rules. This makes the entire process inspectable: you can see which state ran, what data it produced, and why the next transition fired.

---

Key ideas

• States — the named steps in a workflow. Each state has a kind: Agent, System, Human, ParallelAgents, ContainerRun, ParallelContainerRun, or Subworkflow.
• Blackboard — the shared context object that carries data between states.
• Transitions — conditional rules that decide which state runs next based on blackboard values.
• Output template — an optional schema-validated projection of the blackboard, returned as the workflow's final output.
• Scope — controls who can discover and run a workflow: Global, Tenant, or User.

---

Learn more

• Workflows Overview — the conceptual tour
• Building Workflows — manifest syntax and step kinds
• Swarms — multi-agent parallel coordination
• Human Approvals — pause for operator decisions
• CI/CD Pipelines — run container-based build and deploy steps
• Workflow Visualization — render and debug FSM graphs