orchestration-workflow

Multi-agent orchestration manages the spawning and coordination of parallel agents—search, codebase analysis, code execution, and task management—across Plan and Execute modes. The workflow decides when to activate each agent, enforces access controls (blocking direct edits and searches), and collects results for decision-driven branching, reducing manual coordination overhead.

Plan mode presents a structured approach before execution; Execute mode runs it immediately. Orchestration-workflow handles mode switching by reading task files and config, spawning agents conditionally, managing parallel Task tool invocations, and applying result-driven branching. This separation lets you review strategy before committing resources.

The workflow manages initialization, parallel agent spawning, phase execution, and archival by mapping agents to tasks, invoking them in parallel, collecting outputs, and applying retry logic for builds. Result collection feeds branching logic, including USER_INPUT_REQUIRED handling, so downstream actions adapt to what agents return.

Yes. The workflow enforces a unified orchestration layer covering build/test phases with retry logic, commit generation, test-case creation, and archive integration. It manages the full lifecycle—from initialization through execution to result storage—while maintaining access controls and phase sequencing.

Direct search, code edits, and document writing are prohibited at the orchestration layer; agents handle those tasks instead. The workflow enforces read-only access to config.json and task files, manages parallel spawning limits implicitly through Task tool design, and requires USER_INPUT_REQUIRED handling for decisions that need human input.