Use multiple Agents in parallel

Goal

Give a divisible project task to the main Agent, also called the root Agent in conceptual documentation. Have it create subagents to investigate in parallel, then synthesize evidence-backed conclusions in the main conversation.

Before you begin

You must use a project that has already been added. Subagent templates are isolated by project.
Enable at least one available model.
Choose subtasks that can make genuine independent progress, such as “frontend entry point,” “backend data flow,” and “test coverage.” Sequential steps with strong dependencies should not be forced into parallel work.

Steps

1. Create subagent templates

Open Settings → Subagents, select the target project, and choose New template. Enter:

Name: for example, “Frontend reviewer”;
Short description: tell the main Agent when to select it;
Instructions: define its responsibilities, output format, and prohibited actions;
Model: select an enabled model.

Save the template and make sure it is Enabled. Template changes affect only subagents created in the future. Existing nodes retain the snapshot taken when they were created.

You can add separate “Backend reviewer” and “Test reviewer” templates. The more focused each template is, the easier it is for the main Agent to delegate correctly.

2. Explicitly request collaboration in the main conversation

Create a new project conversation and send:

text

Analyze the complete “user saves settings” path. Use subagents to investigate these areas in parallel:
- one for frontend interaction and state;
- one for backend persistence and validation;
- one for test coverage and gaps.

Coordinate the work and perform the final cross-check yourself. Subagents must perform read-only analysis and must not modify files. The final answer must distinguish evidence from the three investigations, shared conclusions, and conflicts, and cite the key files.

The main Agent decides whether and how to call collaboration tools. Explicitly requesting “use subagents” and listing independent workstreams makes it less likely to fall back to sequential single-Agent processing.

Open Subagents in the right sidebar. The list distinguishes active and finished subagents. Open one to view its conversation and status in read-only mode. Users cannot send messages directly in a subagent conversation or start a new turn there.

If a subagent requests approval for a file, command, or another action, handle it from the main conversation. Review the action itself before approving it; do not trust it by default merely because it came from a subagent.

4. Ask the main Agent to synthesize

After the main Agent collects the results, check whether its final answer resolves contradictory conclusions. Ask it to rework or follow up on an item if necessary. A subagent is not deleted after completing one task, so the main Agent can still send it follow-up work.

Expected result

Several subagents with clear task names appear under Subagents in the right sidebar, and the main conversation receives a synthesized conclusion. The original investigations remain available for read-only review. The final answer is compared and reconciled by the main Agent instead of being a simple concatenation.

Troubleshooting

No templates are available: Confirm that a project is selected, the template is enabled, and its assigned model remains available.
The main Agent does not delegate in parallel: State the independent subtasks and “use subagents” more explicitly. The model may still decide that a task is too small to parallelize.
Tasks block one another: Redefine the boundaries so each subtask has independent inputs and deliverables.
A subagent is waiting for approval: Handle approvals only in the main conversation. The subagent view is read-only.
It shows “Outcome unknown”: Do not let the system automatically repeat a step that may have external side effects. Check the actual files or external state first.
There are too many subagents: Reduce the number of workstreams. The system has concurrency and tree-size limits, and more parallel workers do not produce a linear increase in speed.

For the underlying collaboration model, see Multi-Agent principles. For task decomposition, see Manage large projects.