dispatching-parallel-agents

To run independent tasks in parallel without interference, dispatch each task to a specialized agent with isolated context. This prevents cross-task leakage by ensuring agents do not share session history or access each other's data.

Dispatch parallel agents for multi-tasking when you have 3 or more unrelated problems or failures. These tasks must be independent, capable of concurrent investigation, and require no shared state between domains.

To aggregate results from parallel agents running concurrently, implement a coordination layer. This layer collects and consolidates the separate outputs produced by each isolated agent after their parallel execution completes.

No, you do not need shared state to coordinate unrelated failures concurrently. This approach requires strict context boundaries, meaning agents operate in isolation with self-contained prompts and do not reuse shared session history.

The best way to isolate context for specialized agents is to provide each with a self-contained, per-domain prompt. This strict context boundary ensures agents operate independently and prevents cross-task data leakage.

Cross-task leakage happens when investigating independent problems because agents reuse session history or access shared state. Prevent this by dispatching tasks to specialized agents with strict context boundaries and isolated contexts.