Skip to main content

Overview

Workflows let you wire sub-agents into a directed graph that runs multi-step work deterministically — fan out tasks, evaluate results, loop on revisions, and pause for human input. Build and edit them visually in Settings → Extensions → Workflows.

Templates

Start from a template instead of a blank canvas:

Role Agents

Workflow nodes can use purpose-built role agents with tools scoped to their job, so the boundary is physical — a reviewer literally has no edit tools and cannot implement even if asked.
These role agents are workflow-only — they don’t appear in the @-mention picker or the main agent’s dispatch. You can also use your own sub-agents as nodes.

Node Kinds

Edges & Routing

Edges connect nodes and carry a condition: This lets you build pass/revise loops — an evaluator sends work back to a worker until it passes.

Deterministic Checks

A node can run a tool check — a named tool (for example, test_app) that must succeed for the node to pass. If the tool errors, the check fails. This grounds a workflow in real signals like a passing build or test run, not just model opinion.

Breakpoints (Human-in-the-loop)

Building & Running

1

Create a workflow

In Settings → Extensions → Workflows, add nodes (workers, evaluators, joins, managers) and connect them with edges.
2

Add checks and breakpoints

Attach deterministic tool checks where you need hard gates, and breakpoints where you want to steer.
3

Run it

Start the workflow from the context picker in chat. It executes the graph and reports progress.
4

Inspect and steer the run

The run inspector is graph-first: it shows the workflow graph, then per-node results, verdicts, and metrics. While a workflow runs you can pause, step, or terminate it on demand.
Start simple: a worker → evaluator → (revise loop back to worker / pass to join) covers a lot of ground before you add managers and parallel branches.

Sub-agents

The agents that power workflow nodes.

Tools

Deterministic checks call the same tools the agent uses.