Dispatching Parallel Agents

Dispatch parallel agents by identifying independent problem domains—one agent per domain—and execute them concurrently with no shared state. Each agent focuses on a single failure, producing concise summaries that you integrate after completion, significantly reducing investigation time.

Use parallel agents when you have multiple independent failures across different files or subsystems. Sequential debugging wastes time; parallel dispatch lets each agent focus on its own domain simultaneously, cutting total investigation time and keeping scope tightly constrained.

Each agent task must be focused and self-contained with explicit constraints: specify the exact file or subsystem, define the expected output structure, provide detailed context without overlap, and avoid broad directives like 'fix all tests.' Constraints prevent agents from conflicting and ensure independent execution.

Provide a detailed, self-contained prompt template for each agent that specifies the file to investigate, the exact failures to address, constraints on scope, and required output format. Each prompt should be independent—agents must not share state or require coordination to execute in parallel.

No. Parallel agents work only for independent failures with no shared state. If failures are coupled across subsystems or require coordinated changes, sequential debugging or unified investigation is more appropriate to avoid conflicts and data inconsistency.

Each agent returns a concise summary containing the root cause of its assigned failure and the specific changes made. These summaries are structured for easy integration, allowing you to review findings and apply changes without further coordination or consolidation work.