codex-autoresearch

Subagent orchestration in Codex is automated by invoking the autoresearch skill to coordinate a central loop that manages context gathering, hypothesis testing, and mechanical verification. This reduces manual orchestration by maintaining a standing pool of subagents under a central orchestrator.

A subagent-first autoresearch loop is a deterministic workflow where a central orchestrator coordinates subagents to gather evidence, critique, and verify changes. You need this process for large codebases, long-running optimization tasks, and complex experiments requiring parallel analysis and automated logging.

Yes, you can use this autonomous iteration workflow for large codebase optimization. The central orchestrator coordinates plan, verify, and log phases with explicit results per iteration, driving disciplined mechanical iteration toward a measurable goal using a standing pool of subagents.

Yes, you need a root skill bundle at the repository root containing SKILL.md, references, scripts, and agents directories. This structure provides the required tooling to spawn, log, and guide iterations with a codified subagent pool.

The central orchestrator coordinates the plan, verify, and log phases by driving disciplined, mechanical iteration toward a measurable goal. It spawns and guides subagents from a codified pool to handle context gathering, hypothesis generation, and automated verification, logging explicit results per iteration.