Orchestrate multi-agent systems with coordinator-subagent patterns

Let subagents call each other directly — it removes a hop and the coordinator is a bottleneck.

Direct edges destroy the one place that can observe the run, handle errors consistently and control what information moves where. Failures become unattributable.

Hub-and-spoke: every message crosses the coordinator, which is what buys observability and uniform error handling.

Assume a subagent can see the conversation that spawned it.

Subagent context is isolated. Anything not written into its prompt does not exist for it, so it answers confidently from nothing.

Inherited context is a thing you pass, never a thing you get.

Route every request through all subagents so nothing is ever missed.

A one-line lookup then pays the latency and token cost of the whole fleet, and the extra output dilutes the synthesis.

The coordinator's job includes deciding which subagents a query actually needs.

Split a broad topic into one narrow subtopic per subagent and synthesise once.

Narrow partitions leave the space between them uncovered, and a single pass has no way to notice. The report looks complete and is not.

Partition to minimise duplication, then evaluate the synthesis for gaps and re-delegate until coverage holds.

Treat synthesis as the last step — once it has run, the work is done.

That makes coverage a property of the initial decomposition, which was written before anything had been searched.

Synthesis is an evaluation point. Feed its gaps back into targeted re-delegation and re-synthesise.