parallel-execution

Parallel agent workflows reduce end-to-end latency by emitting multiple independent Task tool calls within a single assistant message so they execute concurrently instead of sequentially. You then synthesize the combined outputs into a unified result.

Analyzing multiple codebase directories simultaneously requires issuing all independent Task calls in one assistant message. This ensures concurrent execution across directory layers like authentication, API, and database, followed by a planned synthesis step to merge findings.

TodoWrite tracking coordinates concurrent tasks by marking all parallel todos as in_progress before launching them in a single message. As each Task call completes and returns results, the corresponding todo state transitions are updated to reflect accurate progress.

You should avoid parallel execution when subtasks depend on each other's outputs or modify the same files and shared state. Parallelism requires complete task independence to prevent conflicts and ensure the final synthesis step produces a coherent unified result.

Multi-perspective analysis supports conflicting code review results by launching independent reviews concurrently and resolving contradictions during the planned synthesis step. This combines multiple viewpoints into one unified outcome rather than sequential isolated assessments.

No specific dependencies are required to coordinate independent subtasks concurrently. The approach relies on structuring multiple Task tool calls inside a single assistant message to achieve parallelism, leveraging standard agent workflow components without external modules.