Once a task gets complex enough, 'one super-agent does everything' hits three walls at once: the context window, focus (too many responsibilities in one prompt and the model drops the ball), and maintainability. Codex's direction is Multi-agent: split the task across specialized agents, orchestrate their division and ordering, and pass intermediate artifacts between them via handoffs. In the Agents SDK, Orchestration and Sandbox agents sit as paired standalone modules, signaling that 'orchestration' and 'isolated execution' go together — each sub-agent can run in its own sandbox without cross-contamination. This 'split + orchestrate + handoff' combination is the mainstream path for complex agent systems.
For Zero to Pro Agent Harness learners, Multi-agent and Orchestration are the convergence point of the whole course: the tool calling, permissions, memory, and observability learned earlier all serve 'how to organize a complex task into multiple independently-evolvable, mutually-handing-off agents'. The course repeats a key caution: more splits is not always better — handoffs have real cost (context copying, state alignment), so 'where to split and into how many' is a more important design judgment than 'which framework'. For Zero to Codex learners, the lesson is restraint: get one focused agent working end-to-end before considering multi-agent — most tasks are well served by one agent, and premature multi-agent just adds handoff overhead.