ls-max-swarm

To run parallel coding agents, you use an agent orchestration swarm that decomposes a single large request into two to five independent subtasks, spawning concurrent background sessions via k_bash to execute them simultaneously without blocking.

Agent swarm decomposition is a parallel coding orchestration technique that breaks a single large request into two to five independent subtasks, allowing multiple agents to work concurrently on separate modules for refactoring, documentation, or test-writing workflows.

Yes, you can execute autonomous code review agents without manual approval by using the unattended autonomy flag, which enables optional permission skipping to allow background sub-agents to proceed independently during swarm execution.

Yes, parallel agent orchestration supports read-only exploration for codebase analysis through a specific safety mode flag, restricting the spawned concurrent agents to exploration tasks without modifying the underlying project files.

A key limitation of parallel coding swarms is that task decomposition is appropriate only when a request can be split into two to five independent parts; tightly coupled modules that require sequential execution will block concurrent progress.

To monitor concurrent background coding agent sessions, you track the session IDs reported by the orchestrator immediately after spawning all sub-agents in one message using background PTY execution via k_bash.