OpenAI Ships an Official Claude Code Plugin — and the Multi-Agent Question Flips

Multi-agent coding stops being a theoretical debate and becomes a configuration problem: the plugin makes it cheap to wire up a second opinion, but the real cost shifts to deciding who owns context, who gets write access, and who verifies the final diff. Get the roles wrong and you just bury three windows inside one.

OpenAI deprecated its Codex MCP server and released an official Claude Code plugin under its own GitHub organization. The plugin exposes commands for read-only review, adversarial review, task rescue, status polling, and result retrieval — deliberately keeping review and execution as separate paths. Codex can inspect code and surface problems without touching files; writable tasks require an explicit scope handoff. The design pushes a workflow where one primary agent holds full project context and a second agent is called only for bounded, independently verifiable jobs like diff auditing, stuck-test diagnosis, or small isolated changes. The update does not merge agent brains; it forces a decision about which tool drives and which tool assists. Even with fewer windows, four costs remain: context transfer gaps, permission boundaries, duplicated token spend, and the human acceptance step that no automation removes. A six-item handoff checklist — problem, read scope, write scope, forbidden zones, required verification, and evidence to return — turns out to be the practical interface for making multi-agent delegation work without chaos.