The canonical multi-agent shape. One agent decomposes and synthesizes; the others run bounded subtasks in parallel and never talk to each other.
Why this, for you: orchestrator-worker is the pattern you'll reach for most. But the two halves that
actually determine whether it pays off — decomposition at the front and synthesis at
the back — are the two most people get wrong. The middle (parallel workers) is the easy part.
You may have met the bare shape already. Here we go deeper: the orchestrator's prompt is the
highest-leverage component in the whole system, and synthesis is a reasoning step, not a concat().
1 Two roles, one coordination point
The pattern has exactly two roles, and a single channel between them:
Orchestrator — receives the task, analyzes its structure, decomposes it into independent subtasks, assigns each to a worker, and synthesizes results. It does not execute subtasks itself.
Workers — each receives a bounded, self-contained subtask with its own scoped tool set, explores independently, and returns results.
Workers returning results to the orchestrator is the only coordination point. Any
state sharing between workers during execution is a design smell — it creates coupling that undermines the
parallelism you went multi-agent to get.
2 Decomposition: scale effort to complexity
The orchestrator should match worker count and tool allocation to task complexity. Anthropic's research system
documents explicit effort-scaling rules — and these belong in the orchestrator's system prompt, not
hard-coded in the harness:
Query
Workers
Tooling
Simple
1 agent
3–10 tool calls
Moderate
2–4 subagents
clearly divided responsibilities
Complex
10+ subagents
carefully partitioned search spaces
Hard-coding a fixed agent count removes the flexibility to match scale to complexity — and over-spawning workers on
simple queries is the most common failure of the pattern. The orchestrator needs the freedom to send one
worker when one will do.
The orchestrator prompt is the highest-leverage component
Small changes to the orchestrator's prompt can unpredictably shift how subagents behave — its
decomposition decisions determine which subtasks every worker receives. Test decomposition behavior explicitly
across a range of input queries; one fragile orchestrator routes every worker to the wrong job at once.
3 Synthesis is reasoning, not concatenation
After workers complete, the orchestrator synthesizes. This is the half teams skip — and skipping it turns the
pattern into pure latency tax. Synthesis is a reasoning step where the orchestrator:
Evaluates the reliability of each worker's findings
Identifies conflicts or gaps between results
Produces a unified output drawing on the strongest elements from each worker
If the orchestrator simply concatenates worker outputs, the pattern adds latency and
token cost without improving quality. The reasoning at synthesis is what justifies the architecture.
And synthesis has a hard ceiling: the orchestrator must hold the task plus every worker's results. Beyond
roughly 4+ substantive outputs this routinely overflows practical context budgets — so workers should
return distilled findings, not raw transcripts. (More on that contract in Part 2.)
# Orchestrator system prompt — effort scaling lives HERE, not in code"You receive a list of repository paths. For each path, spawn a worker with tool access limited to that repo. Workers run in parallel. When all workers return, SYNTHESIZE into a ranked list by severity — resolve conflicts, do not concatenate raw findings."
↪ Your win: front and back, not just the middle
Two roles, one channel — workers report to the orchestrator and never to each other.
Put effort-scaling rules in the orchestrator prompt — 1 / 2–4 / 10+ workers by complexity, not a hard-coded count.
Treat the orchestrator prompt as highest-leverage — test its decomposition across varied inputs.
Synthesize, don't concatenate — evaluate, reconcile conflicts, draw the strongest elements together.
Watch the synthesis ceiling — 4+ substantive worker outputs overflows context; distill before returning.
Retrieval practice — recall, don't peek
Question 1The only legitimate coordination point in the pattern is…
Question 3If the orchestrator merely concatenates worker outputs, it adds…
Question 4The highest-leverage component, where small changes cascade, is the…
Question 5 · spaced recall from Lesson 1You should climb to the multi-agent rung only when…
Ask me anything. Want a worked orchestrator prompt for a real decomposition of yours, or how to
keep synthesis under the context ceiling? Next in Part 1: Fan-Out and Synthesis — N agents on the
same problem, and how to merge them.