swarm

Parallel execution without a lead agent uses a swarm model where workers self-coordinate through shared state, claim tasks, and run concurrently up to a configured limit to maximize throughput across wide dependency graphs.

Parallel execution tracks task completion and automatically unblocks downstream dependencies, dynamically re-queueing ready tasks into a defined ready-set so available workers can resume processing without waiting for a central lead.

Yes, you can isolate concurrent file operations using optional worktree isolation, which creates safe execution environments for parallel workers to operate without file conflicts when processing largely independent tasks.

The best way to coordinate independent parallel tasks is a swarm model, allowing workers to self-coordinate through shared state rather than relying on a lead agent, enabling high-throughput processing when file conflicts are rare.

Failure handling in parallel execution includes robust mechanisms like deadlock detection, identification of zombie tasks, and safe merge policies, ensuring the system safely recovers from stalled or failed worker lifecycles.

Avoid a swarm model for parallel execution when tasks have frequent file conflicts or tight interdependencies, as the model is optimized for largely independent tasks where dynamic blocking and re-scheduling overhead is minimal.