STEP 6 OF 8

Workflows

Agent Studio (sidebar label; route is /workflows) is the drag-and-drop, node-based orchestration canvas. /workflows lists what you have and offers two starting points: describe it (a natural-language first draft) or a blank canvas. /workflows/{id} is the ReactFlow builder itself.

Starting from a blank canvas or a description

# Natural-language -> first-draft nodes/edges
curl -X POST https://your-control-plane-host/api/v1/workflows/ai/generate \
  -H "Authorization: Bearer $TOKEN" -H "Content-Type: application/json" \
  -d '{"description": "When a new invoice document lands in the Vault, classify it, route amounts over $5,000 to a human for approval, then post the result to Slack."}'

Or create an empty one and build it node by node from the palette:

curl -X POST "https://your-control-plane-host/api/v1/workflows?tenant_id=$TENANT_ID" \
  -H "Authorization: Bearer $TOKEN" -H "Content-Type: application/json" \
  -d '{"name": "Invoice approval", "definition": {"nodes": [], "edges": [], "viewport": {}}}'

The definition field is stored exactly in ReactFlow's shape ({nodes, edges, viewport}) — the canvas reads and writes it directly, no translation layer.

The real node-type palette

The palette isn't hardcoded on the frontend — it's fetched from the backend, so it always matches what the executor actually supports:

curl https://your-control-plane-host/api/v1/workflows/node-types

The current set (app/api/v1/endpoints/workflows.py::get_node_types, dispatched by app/services/workflow_executor.py):

| id | Name | Purpose | |----|------|---------| | trigger | Trigger | Entry point for the workflow | | agent | Agent | Calls a workforce agent (the real palette id — matches _handle_ai_agent in the executor) | | classify | Classify | Model-based classification step | | condition | Condition | Branches on a condition (2 outputs) | | while | While loop | Loops while a condition holds | | transform | Transform | Reshapes step output | | set_state | Set state | Writes to workflow state | | parallel / merge | Parallel / Merge | Fan-out to N branches / fan back in | | integration | Integration | Calls a connector action | | mcp | MCP tool | Calls an MCP-exposed tool | | human_review | User approval | Pauses for a human decision (2 outputs); routes through Action Gateway pause/resume | | guardrail | Guardrail | Policy check gate (2 outputs) | | file_search | File Search | Searches ingested documents | | output | Output | Terminal node |

Drag a node from the palette, connect handles, then use the side panel to edit the selected step's label and instructions.

Save, test, and activate

# Dry-run without side effects
curl -X POST https://your-control-plane-host/api/v1/workflows/{workflow_id}/simulate \
  -H "Authorization: Bearer $TOKEN"

# Real run -- deducts credits
curl -X POST https://your-control-plane-host/api/v1/workflows/{workflow_id}/execute \
  -H "Authorization: Bearer $TOKEN"

# Flip from draft to active
curl -X POST https://your-control-plane-host/api/v1/workflows/{workflow_id}/activate \
  -H "Authorization: Bearer $TOKEN"

GET /api/v1/workflows/{workflow_id}/executions lists run history; schedules and webhooks (POST .../schedules, POST .../webhooks) trigger a workflow on a cron or an inbound event instead of a manual execute call.

Demo sessions carry no real backend JWT, so demo users see two grounded example workflows (AP invoice approval, compliance evidence collection) loaded from lib/demo-data/workflows.ts. Save/Test run/Activate work against that local state only — nothing above hits a real backend for a demo account.

Starting from a Solution Template instead

Complex Use-Case (sidebar label; route is /solutions) is a library of pre-built, pre-configured templates — each one ships with pre-assigned agents, connectors, and a runnable workflow already wired together. Pick a template, then configure it with natural language or voice before deploying:

# Pull one template's full pack (agents, connectors, workflow) for review
curl https://your-control-plane-host/api/v1/solutions/{use_case_key}/pack \
  -H "Authorization: Bearer $TOKEN"

# Deploy it -- provisions the real agents, skills, and workflow for your tenant
curl -X POST https://your-control-plane-host/api/v1/solutions/{use_case_key}/provision \
  -H "Authorization: Bearer $TOKEN"

The same endpoints are also mounted at /api/v1/solution-templates/{use_case_key}/... — an alias with identical behavior, not a different implementation.

What's next

Step 7 — Evals & Monitoring: once agents and workflows are running, watch them via traces, threads, and eval regressions.

← Step 5 — Skills & Code

Step 7 — Evals & Monitoring →