Orchestration wires several agents into one scenario — a graph where the
result of one step decides the next. It is how you build a process (for example:
classifier → doer → reviewer) instead of relying on a single large prompt. The
screen is workspace-scoped.
Visual editor
The primary editor is a visual flow builder in the make.com style: draggable
node cards, SVG connection edges, an add-node palette and a side panel to
configure the selected node. Node positions are saved with the graph, so the
diagram looks the same every time you reopen it.
What a graph is made of
A graph is data: { start, max_steps, nodes:[…] }. Each node holds an id, a
bound agent, a prompt, a retry count max_retries and a list of edges.
An edge condition (when) can be: always, default, contains /
not_contains / equals / regex (the last four need a value). The reserved
terminal node is __end__.
How to use it
Click “+ Create”, set name, description and enabled.
“Insert example” loads a ready graph template — a handy starting point.
Add nodes, pick an agent and prompt for each, and connect them with
conditional edges.
Validate — check the graph before running.
For an existing graph, enter a dry-run input and click Run — you get a
step-by-step trace: each node’s output, agent, errors, stop reason and the
final result.
Export and import
The “Export/Import JSON” toggle exposes the graph’s raw JSON in a textarea
with a live human-readable preview — convenient for copying, versioning and
moving graphs between instances.
Configuration
max_steps caps the total number of steps (loop protection).
A node’s max_retries sets how often a step is retried on error.
The graph’s enabled flag decides whether it can be run.
**Orchestration** wires several agents into one scenario — a graph where the
result of one step decides the next. It is how you build a process (for example:
classifier → doer → reviewer) instead of relying on a single large prompt. The
screen is workspace-scoped.
## Visual editor
The primary editor is a **visual flow builder** in the make.com style: draggable
node cards, SVG connection edges, an add-node palette and a side panel to
configure the selected node. Node positions are saved with the graph, so the
diagram looks the same every time you reopen it.
## What a graph is made of
A graph is data: `{ start, max_steps, nodes:[…] }`. Each **node** holds an `id`, a
bound `agent`, a `prompt`, a retry count `max_retries` and a list of **edges**.
An edge condition (`when`) can be: `always`, `default`, `contains` /
`not_contains` / `equals` / `regex` (the last four need a value). The reserved
terminal node is `__end__`.
## How to use it
1. Click **"+ Create"**, set name, description and enabled.
2. **"Insert example"** loads a ready graph template — a handy starting point.
3. Add nodes, pick an agent and prompt for each, and connect them with
conditional edges.
4. **Validate** — check the graph before running.
5. For an existing graph, enter a dry-run input and click **Run** — you get a
step-by-step trace: each node's output, agent, errors, stop reason and the
final result.
## Export and import
The **"Export/Import JSON"** toggle exposes the graph's raw JSON in a textarea
with a live human-readable preview — convenient for copying, versioning and
moving graphs between instances.
## Configuration
- `max_steps` caps the total number of steps (loop protection).
- A node's `max_retries` sets how often a step is retried on error.
- The graph's enabled flag decides whether it can be run.
## Next
- [Agents](/en/v1.0/webui/agents) — the graph's participants.
- [Orchestration & subagents](/en/v1.0/concepts/orchestration-subagents) — how it
works under the hood.
- [Sessions](/en/v1.0/webui/sessions) — where nested subagent runs are shown.
