The practice of assembling several specialized agent roles into a pipeline or workflow graph, coordinated by an orchestrator that routes work between them. The top layer of the AI-Native Developer Workflow, built on Loop Engineering.
As the workflow scales, the human's role shifts from writing prompts directly to designing a system of roles — mirroring how a real product team distributes responsibility:
A main "orchestrator" session pulls an issue from the backlog, routes it to the PM agent if not yet groomed, then to the engineer to implement, then to QA to pass/fail, looping back to the engineer on failure.
The QA/tester role is argued to be genuinely valuable rather than process theater: it is reported to catch mistakes the implementing agent missed, echoing the rationale for human code review that self-review is less reliable than independent review.
The main agent session can be designated an orchestrator: instead of grooming, implementing, or testing itself, it only launches and sequences sub-agents for each role (PM, engineer, QA — see agent-team-personas) and hands off between them. The orchestrator is explicitly instructed not to do the work directly, only to route it.
This depends on sub-agents — a capability the presenter reports as supported by Codex, Claude Code, and likely other engines — letting a main session spawn separate, isolated agent sessions per role. Sub-agents are the concrete mechanism that makes the orchestrator pattern possible; without them the main session would have to context-switch personas in place instead of delegating.