orchestrator

Coordinate autonomous research by maintaining hypotheses in RESEARCH_STATE.md, designing experiments, and delegating execution to specialized subagents. This approach manages planning and synthesis while restricting model invocation to scale inquiry without manual micromanagement.

Starting subagent coordination requires a RESEARCH_STATE.md file for current thinking, a tools directory with reusable utilities, and hooks to control subagent behavior. You initialize this project structure and load core skills to spawn a scientist for the first experiment.

Yes, spawn and coordinate specialized subagents like scientist, colleague, and judge within a single workflow. The orchestrator manages these agents under constrained access via hooks to guide behavior during experiments and synthesis.

Experiment design and execution are handled by loading supporting skills like experiment-structure at session start, planning the experiment, then spawning scientist subagents to execute it. The orchestrator delegates this execution while maintaining the overall research state.

Restricting model invocation limits subagent autonomy by controlling behavior through hooks. This ensures subagents operate within constrained access boundaries but requires careful hook configuration to prevent stalling during delegated experiment execution and research synthesis.