Agent orchestration is the operating layer between specialised workers. The test is not how many nodes the canvas shows. It is whether ownership, handoff, review, and state remain visible across each boundary.
What is agent orchestration? The assignment and coordination of work across human and AI workers so the next owner can act from a complete packet, pass a written gate, and update shared state.
Why multi-agent systems fail owns the coordination gap. This page owns the boundary test.
The four-part boundary test
Ask four questions.
- Who owns the work now, and who owns it after the boundary?
- What packet crosses with the work?
- Which written criteria can fail it before downstream use?
- Where can every responsible worker see current state?
If any answer is missing, the canvas is showing activity rather than coordination. What is an AI agent? owns the worker. Orchestration is the operating layer between workers.
A runtime graph can encode serial work, fan-out, fan-in, and escalation. That is useful routing. It becomes coordination only when the owner, packet, gate, and state surface are also explicit.
Collision control
CHORUS is one implementation. CHORUS owns the named six-skill protocol. This page owns the ordinary operating term.
Not “we added a second agent.” Worker count says nothing about ownership, review, or shared state.
Not the definition of an agent. An agent holds a goal and uses tools. Orchestration determines how several workers share work without losing accountability.
What this page is not
It is not the six-skill expansion and not a vendor bake-off. It is the four-part boundary test.
The Multi-Agent Organization is live. Write the four questions on a card. If the answers are empty, stop calling the graph orchestration.