Codex CLI /agents: Stop Driving One AI, Start Running an AI Team

The original author is Kai Good (@kaigooddev), and this is a faithful rendering of his hands-on notes posted on X on October 7, 2026. It covers the new /agents view in Codex CLI, splitting large work across agents in parallel, and the conflict problem he openly leaves unresolved.

The original note: don't give everything to one agent

Working with a single Codex is starting to feel like the old way. The new Codex CLI has "/agents". Instead of giving everything to one agent, you split work across several agents and watch their progress on one screen.

The point is simple: for large jobs, breaking work into small pieces and running them in parallel beats pushing one agent through everything. The author says he now works with role-specific agents under one orchestrator.

Lately I've been putting role-specific agents under an orchestrator, and for big jobs, splitting things up and running them in parallel clearly works better than hammering one agent nonstop.

Then comes the line that defines the piece: as agents multiply, his own role shifts from developer to team lead.

The problem is, the more agents I add, the more I turn from a developer into a manager lol

He sums up the shift as vibe coding moving "from making AI code to running an AI team." He also shared the official new Codex CLI announcement link in the same thread.

The still-open question: who stops two agents touching the same file

When asked how he prevents agents from editing the same file at the same time, and whether he gives each agent its own branch, the author answers honestly that he has no fixed method yet.

I don't have a fixed approach for that yet. I still need to experiment more on whether the rule of never assigning the same file at the same time or branch isolation is the right answer.

So the only two confirmed candidates in this tip are "the rule of never assigning the same file simultaneously" and "branch isolation" — and even the author is still experimenting. No fixed commands, config examples, or branch strategy are added here because none exist in the source.

Background: the confirmed scope of the subagent workflow

The corroborating material only supports this much: the Codex subagent workflow has a manager agent split work across specialist agents in parallel and collect the results into one response, spawning subagents only when the user explicitly requests it. Agents can carry their own model, sandbox, and MCP settings, while subagents inherit the parent's sandbox policy and cannot escalate above it. The figures of a default max of 6 concurrent threads and depth 1 appear in the same confirmed material. That is background from public docs and coverage, not the author's own workflow, so actual behavior per version and plan should be checked against the official announcement.

Original source

• Kai Good (@kaigooddev) original thread: https://x.com/kaigooddev/status/2107753518732054704
• Official new Codex CLI announcement shared by the author: https://openai.com/ko-KR/index/devday-2026-recap/