Skip to content

Workflows

The Studio's Workflows view is a drag-and-drop canvas (AutoGen-Studio "Team Builder" style) for composing a multi-agent workflow, then running it — with zero graph-engine changes.

The mapping

  • A node is a step assigned to a role (an instruction + a role from the catalog).
  • An edge is a dependency.
  • The canvas is a DAG run as a swarm: independent nodes run in parallel waves, dependent nodes run after their upstreams, and results are shared through the blackboard.

A tiny StaticPlanComposer replays the hand-drawn graph: node → subtask, role-per-node, edge → dependency. The existing orchestrator honors a composer-supplied plan, so the user's graph runs for real.

Run-level controls

Model and temperature are run-level controls on the canvas toolbar (the swarm worker uses one global worker_model/sampling). Per-node tool scope comes from the chosen role. True per-node model/tool overrides are a future engine change.

Endpoints

Endpoint Purpose
GET /api/workflows · POST /api/workflows list / save (upsert) named workflows
GET /api/workflows/{name} · DELETE /api/workflows/{name} load / delete
POST /api/workflows/run run a spec → RunResult
GET /api/workflows/run/stream run with a live SSE node trace

A cycle, duplicate id, dangling edge, or empty graph is rejected with HTTP 422.

Library use

from riptide_watergraph.workflows import WorkflowSpec, WorkflowNode, WorkflowEdge
from riptide_watergraph.service import run_workflow

spec = WorkflowSpec(
    name="research-then-summarize",
    nodes=[WorkflowNode(id="a", role="researcher", subtask="research water"),
           WorkflowNode(id="b", role="scribe", subtask="summarize the findings")],
    edges=[WorkflowEdge(source="a", target="b")],
)
result = run_workflow(spec, offline=True)
print(result.mode, result.results)