🗺️ Orchestration patterns

Three patterns cover 90% of real-world cases: independent parallel work (all at once, with no dependencies), pipeline (A hands off to B, who hands off to C) and fan-out (one launches many that converge on a result). Understand the differences between them — especially the hidden cost of agent teams — is what separates those who designs systems of someone who only uses subagents.

↑ Three topologies: each has different use cases, costs, and latency.

The simplest pattern, and the most powerful when jobs don't depend on one another. You have N tasks that can run at the same time, each in its own subagent, and the conductor collects the results at the end. No subagent needs to wait for another—the total time is the slower, not the sum of all of them.

When one agent’s result is the input for the next one, you have a pipeline. There’s no parallelism here — the flow is sequential by necessity. The benefit isn’t speed: it’s specialization. Each stage has an agent focused only on that, with its own tools and instructions. Research is different from drafting; drafting is different from reviewing.

Fan-out is parallelism with explicit convergence: the conductor launches N subagents and then synthesizes their reports into a single result. This is the “5 personas demo” pattern—each subagent uses a different lens, but they all report back to the same conductor, who assembles the final view.

Agent teams are when multiple agents talk to each other — each message from one agent to another costs tokens. A regular subagent doesn't talk: it receives a brief, works in isolation, and delivers a report. In agent teams, the conversation itself has a cost — and this cost can exceed the value delivered if it isn't controlled.

There is a turning point where continuing in the same context becomes counterproductive. We call it context wall: the point at which the model starts losing earlier information, mixing data from different files, or getting confused by cross-references. The rule of thumb: when you cross the 10 files, consider splitting it up.

The 15-chapter example is the best practical illustration of the pattern. You have a book with 15 chapters and want to know: "where is the weakest argument?" Doing this in a single context means reading all 15 chapters in sequence — the model starts mixing up the first chapters with the last ones. With 15 subagents in parallel, each person reads a chapter and submits a 1-page report. The conductor reads 15 one-page reports and synthesizes them.

There’s a sequence of three questions to find the right topology. Answer in order—the first “yes” determines the default. If none apply, a single agent is probably enough.

The table below maps the four patterns to concrete use cases. Use it as a quick reference when designing a system. For each case, the “Why?” column gives the deciding argument.

---
name: chapter-auditor
description: Revisa um único capítulo de texto em busca
  de argumentos fracos, inconsistências e lacunas.
  Use quando o maestro precisar auditar um capítulo
  isolado (padrão fan-out de N capítulos).
tools: Read                 # só lê o arquivo do capítulo
model: haiku               # rápido e barato por capítulo
---

You are a critical editor. You receive ONE chapter.
Identify: weakest argument, factual gaps, unclear sections.
Return a structured report: 3 bullets max. Be blunt.

Three prompts to help you identify and use topologies in practice. The pattern: describe the problem, point out the dependency (or lack of one), and ask for the right topology.

Tenho 20 arquivos de código para auditar. Cada um é
independente dos outros. Qual topologia de subagentes
devo usar? Justifique e mostre como estruturar o briefing
do maestro para disparar todos em paralelo.

Use fan-out: dispare 1 subagente por capítulo para revisar
os 15 capítulos em paralelo. Cada sub deve retornar um
relatório de no máximo 3 bullets: argumento mais fraco,
lacuna factual, seção confusa. Sintetize no final.

Monte um pipeline de 3 agentes para criar um artigo:
Agente A: pesquisa e extrai 10 fontes sobre [tema].
Agente B: recebe as fontes de A e escreve o rascunho.
Agente C: recebe o rascunho de B e revisa para publicação.
Use Haiku em A, Sonnet em B e C.

Here's the terminal with a fan-out of 5 chapters underway. The conductor launched 5 subagents at the same time; each one reads 1 chapter with Haiku. Note: the conductor's context is at 9% — all the heavy reading happens in the subagents' windows.

↑ Illustrative recreation of the terminal. 5 chapters in parallel with Haiku; orchestrator in Sonnet, context almost untouched.