One lead, many specialists
You built the pieces across this track: routing a task to the right agent, fanning work out to ordered workers, voting away disagreement, coordinating through a shared board. Orchestration is the pattern that uses all of it at once. A lead agent — the orchestrator — reads a goal, splits it into subtasks, and hands each to the specialist best equipped for it, then stitches their answers into one result.
The power is specialization. No single agent has to be good at everything: the orchestrator isn't a researcher or a calculator, it's a delegator. Below it has a real team — a researcher (it can look things up on the live web and in the handbook), a mathematician (it converts and computes with real tools), and a writer (it turns raw facts into a clean sentence). Each is a genuine agent with its own instructions and its own tools, running for real.
Watch the orchestrator work the question "how tall is the Eiffel Tower, in metres and
feet?": it delegates the lookup to the researcher, sends that number to the
mathematician to convert, hands both figures to the writer, and reports back the
finished sentence. You'll see each ↳ delegating to … handoff and each specialist's
own reasoning and tool calls, labelled, before the orchestrator's final answer.
The failure mode to watch for is the broken hand-off: the researcher replies "around 330 m, give or take a few since the antenna height varies" and the orchestrator forwards that whole sentence to the mathematician, which can't multiply prose by 3.28. A weak orchestrator passes outputs through untouched; a good one treats each specialist's reply as raw material, extracts the one figure the next agent needs, and only then delegates — so the chain stays clean even when a specialist is chatty.
A capable agent team isn't one genius — it's a coordinator who knows who to ask. The orchestrator's only skill is decomposing the goal and routing each part; the specialists do the rest.