Graph & Workflow

Choose orchestration based on the control flow you need, rather than treating every multi-step task as a multi-agent problem.

Need	Abstraction
Sequential stages, parallel steps, routing	workflow.Pipeline, Parallel, Router
Explicit acyclic dependencies	graph.Graph compiled as a DAG
Cycles and iterative shared state	graph.PregelGraph
Homogeneous typed values through nodes	graph.TypedGraph[T]
Delegate to another agent or transfer control	agentcore.HandoffConfig

Continue with Runnables and streams, graph execution, or workflows and handoffs for executable examples and semantics.

Workflows

Pipelines execute ordered steps. Parallel workflows combine independent results through a merge function. Routers select a branch from the current input. Use the workflow API for step contracts and execution options.

Graphs

Build a DAG by adding nodes and edges, then compile it with an entry node. Pregel graphs support repeated super-steps and conditional routing through shared state. Compilation provides diagnostics and deterministic introspection; compiled graphs also expose Mermaid export for architecture documentation.

Runnable[I, O] supports Invoke, Stream, Collect, and Transform modes. Typed node codecs bridge typed runnables into the string-based graph engine; TypedGraph[T] keeps homogeneous values typed without that conversion.

See the graph API and Runnable API.

Agent Handoffs

A handoff target uses AgentConfig to configure a local specialist. HandoffDelegate delegates a task, while HandoffTransfer transfers control. Remote targets can be registered with the A2A integration.

The CLI example configures weather and math specialists with a coordinator, illustrating local handoffs alongside tool execution.