Skip to main content
Workflows are deterministic, versioned DAGs you build on a visual canvas. Drop in a trigger, chain connector actions, branch on conditions, run code, and (if you want) call an AI Agent as a node. Save it and the workflow can be invoked by another agent as a tool — composable automation without leaving the platform.

​
Anatomy

A workflow has four tabs at the top: General, Canvas, Versions, and Alias — mirroring the surfaces you already know from agents.

​
General

​
Canvas

The visual DAG editor. The right rail holds the node palette, grouped into three sections:
Agent traces

​
Triggers

Define how the workflow is started. Each workflow has exactly one trigger.

​
Nodes

The building blocks of the DAG. The connector nodes (Gmail / Jira / Calendar / GitHub / Slack / Telegram) come from connectors — each connector action surfaces as a typed node with typed inputs and outputs.

​
Control

Flow-control primitives.

​
Building the graph

  • Add a node: drag from the palette onto the canvas, or click. Connect the source node’s output handle to the next node’s input handle.
  • Configure a node: double-click it. The right panel opens with the node’s editable fields:
    • ID — used in references (e.g. {{node_4.field}}).
    • Label — display name on the canvas.
    • Connection — for connector nodes, which OAuth connection to act through.
    • Inputs — action-specific parameters. String fields support templating with {{node_id.field}} references to upstream outputs.
    • Outputs — the typed return shape, with copy-to-clipboard buttons for each field reference (e.g. {{node_4.next_page_token}}).
  • Delete a node: select it and press Delete, or use Delete Node at the bottom of the config panel.
  • Save the canvas to persist your changes. Export JSON dumps the workflow definition.
Outputs from any node are referenced downstream as {{node_id.field}}. The ID is editable, so you can rename node_4 to list_messages for readability.

​
Versions

Workflows are versioned with the same model as agents:
  • $LATEST is version 0 — the mutable working copy. Every save on the Canvas tab updates \$LATEST.
  • Create New Version freezes the current \$LATEST into a numbered immutable snapshot (1, 2, 3, …). Future canvas edits go to \$LATEST again; numbered versions never change.
  • The Versions tab lists every snapshot with its creation/update timestamp. Click View to open a version read-only in the Canvas tab.

​
Alias

Aliases pin a named label (e.g. production, canary) to one or two versions:
  • $LATEST is a built-in alias that always points at version 0.
  • Custom aliases can point at a single version, or split traffic across two versions with a weight (for canary or A/B rollouts).
  • Agents that invoke this workflow can reference it by alias (e.g. pr_triage@production) instead of a numeric version, so you can roll forward without touching the agent.

​
Invoking a workflow from an agent

When the Invoked by Agent trigger is used, the workflow registers itself as a tool with the same name as the workflow. To make the agent able to call it:
  1. Open the agent in Agent Builder.
  2. Go to the Tools tab.
  3. Enable Workflows and select this workflow (and the version or alias you want to pin).
  4. Save.
The agent now sees the workflow as a tool. It calls the tool with arguments matching the trigger’s input schema; the workflow runs to completion (durably on Temporal, in-process on Local) and the agent receives the output back as the tool result. This is the workflow-as-tool composition pattern: build complex multi-step automation deterministically as a DAG, then let an agent orchestrate it like any other tool.

​
Local vs Temporal runtime — when to use which

You can change the runtime at any time from the General tab; existing version definitions stay the same.

​
Observability

Every workflow run produces an OpenTelemetry trace covering:
  • The trigger event.
  • Each node executed, with inputs, outputs, status, and duration.
  • Tool calls made by any AI Agent nodes inside the workflow.
  • Re-tries and replays (Temporal).
See Tracing for how to filter and inspect runs.

​
Relation to other features

  • Connectors — provide the typed connector nodes (Gmail, Slack, Jira, …) on the workflow palette.
  • Agent Tool — agents call workflows just like any other tool, with arguments derived from the trigger’s schema.
  • Agent Runtime — workflows mirror the runtime choice agents have: in-process for speed, Temporal for durability.