# Agents spawning agents

Solo agents can delegate work to other Solo agents. The important part is that delegation is not limited to the same agent tool, model, provider, or lab.

A Claude session can ask Solo MCP to spawn a Codex worker. A Codex session can spawn Gemini, Amp, OpenCode, Claude, or any custom agent tool you have configured. The parent agent does not need to come from the same lab as the child agent. If the target agent tool is launchable for the effective project environment and [MCP tools](../mcp-tools/overview.md) are enabled for the caller, the lead agent can choose the right worker for the lane.

For the broader model behind this, read [The agentic metaharness](/blog/the-agentic-metaharness).

## What this unlocks

Use one lead agent as the coordinator, then let it pick specialist workers:

- A planning-heavy lead can spawn implementation workers.
- A code-focused lead can spawn a second agent for docs, tests, or review.
- One lab's agent can ask another lab's agent to check assumptions or compare approaches.
- A custom team agent can spawn a general coding agent for a bounded repo edit.

This makes Solo an orchestration layer instead of a single-agent launcher. The lead agent can inspect the project, list available agent tools, [spawn workers](../mcp-tools/agent-terminal.md), send prompts, track todos, [set timers](../mcp-tools/timers.md), and inspect worker output before integration.

## What to ask the lead agent

Give the lead agent permission to choose tools by fit, not by brand:

`Use Solo MCP to list the available agent tools. Spawn whichever agent is best for each lane, even if it is from a different lab than you. Keep the workers bounded, track them with todos, and report which agent handled each lane.`

For larger work, pair that with a scratchpad and todo plan:

`Interview me, write the plan into a scratchpad, split it into todos, then spawn agents from any available lab for the independent lanes. Set blockers where work depends on another lane, and set a Solo timer to wake you when workers go idle.`

## Keep ownership clear

Give every worker a narrow assignment to keep ownership clear. The lead should tell each child agent what it owns, what it should avoid, what context matters, and what it must report back.

The child agent does not inherit judgment automatically from the parent. Treat each spawned worker as a separate collaborator with a clear prompt, a bounded task, and a concrete handoff.

## Requirements

- The target agent tool must be configured and enabled in Solo.
- The target tool must have a launchable installation for the effective project environment. Ready and unknown-health tools can be launched; missing or broken installations cannot.
- The lead agent must have access to Solo MCP tools.
- The spawned agent runs in the effective Solo [project scope](../mcp-tools/project.md).
- Model, permission, and CLI flag differences may come from the configured agent tool or per-launch `extra_args`. The initial prompt is delivered after launch, so it cannot change launch-time CLI flags; any model or permission change it requests depends on the agent CLI.

See [Agent orchestration workflow](./agent-orchestration.md) for the fuller lead-agent pattern.

---

Are you a human? Read this doc on the web: https://soloterm.com/docs/workflows/agents-spawning-agents