Module 4 · Working Like a Pro
4.1 Multi-Agent Delegation
A single Codex session, everything up to this point, is one agent doing one job at a time, however large that job is. Multi-agent delegation is the step past that: Codex can spin up subtasks and hand them to parallel subagents, each configurable with its own model choice and reasoning level, working on separate pieces of a larger goal at the same time instead of one long sequential thread.
The instinct this mirrors is exactly how you'd actually run a team: a big migration touching three independent services doesn't need one engineer working through them one at a time if three people can take one service each without stepping on each other. Delegate research or exploration — "figure out how this legacy module handles errors" — to a subagent so your main session isn't carrying forty files of exploration context for the rest of a conversation that has nothing else to do with those files. You get the two-paragraph conclusion back, not the journey it took to reach it.
Be honest with yourself about when this is actually warranted, because it's easy to reach for the impressive-sounding option when the simpler one would've been faster. Most day-to-day work — a bug fix, a small feature, a routine refactor — doesn't need delegation at all. It needs one focused session. Multi-agent delegation earns its complexity specifically on work that's genuinely large and genuinely separable: independent pieces of a system that really can move in parallel, or a research task you'd rather receive a summary of than personally wade through.
Do this before moving on
think back to one recent project big enough that you actually wished you could split it across a couple of engineers working in parallel. That's the shape of problem multi-agent delegation is built for — and if nothing comes to mind, that's a fine answer too. Not everything needs this every week.