How a run executes

Understanding how a multi-agent run executes turn by turn helps you write better instructions and build more reliable agents.

The execution model

• The root agent never calls the model itself — it only holds shared instructions and routes between sub-agents. Every real answer comes from a sub-agent.
• The root's shared instructions are merged into whichever sub-agent is active, so they always influence behavior even though the root never "speaks" on its own.
• A sub-agent's own instructions take priority over the root's shared instructions when the two would conflict — write sub-agent instructions assuming they win.
• A run proceeds in steps; each step is one sub-agent's turn (one or more tool calls, a handoff, or a final answer).
• Every run ends in exactly one final answer.

What you see in traces

Routing timeline rows always show a real sub-agent as the responder — never the root agent itself as the "from" or "to" party for a model-generated answer.

Authoring implications

• Put agent-wide policy in root instructions; put anything behavior-critical in sub-agent instructions when the two could conflict.
• Always define at least one sub-agent as the default — a root-only agent with no sub-agents cannot answer anything.