Parallel subagents in one sandbox (Beta)
Agents can use parallel_agents to choose a team size for each task and delegate subtasks concurrently. Each child uses the parent’s
sandbox and model, with its own conversation. The parent supplies the task context, gathers the answers and
can read the children’s files before completing the larger task.
The default concurrency ceiling is eight. Set spec.maxParallelAgents (1–32), or max_parallel_agents= in
Python, to override it for new runs. Ask for a specific team size in the prompt when needed.
Children share the parent’s model budget and turn limits. Completed child steps survive retries, and cancelling the parent stops further rounds and cleans up the shared sandbox. Children cannot delegate recursively.