Home / Several agents and adversarial checking

Fan-out, pipeline, barrier: choosing the architecture, or giving it up

Fan-out launches independent agents in parallel behind a shared barrier, a pipeline lets each item pass through without a barrier and remains the default for multi-step work, a barrier is only justified when a step genuinely needs the full result of the previous one, and sometimes the right decision is to use no agent at all when a single call already suffices.

Three shapes come back every time several agents work on the same task. Fan-out launches independent agents in parallel and brings them together behind a shared barrier, which waits for all of them to finish before moving on. A pipeline lets each item pass from agent to agent without ever stopping to wait for the others, and this is the default shape as soon as a task splits into several steps. The third option, often forgotten, is to launch no agent at all.

When the barrier is justified

A barrier is expensive: it blocks everyone until the last agent arrives, and the slowest result sets the pace for the whole group. It is only justified if the next step genuinely needs the full result of the previous agents to start, comparing three quotes before choosing a supplier is one example. If each subtask can move forward without waiting for the others' results, a pipeline without a barrier moves faster: each item processed moves on to the next step as soon as it is ready, without waiting for the others to reach the same point.

Giving it up, the fourth box

Anthropic documents this multi-agent architecture as an expensive option, not as a flaw. A multi-agent system consumes, by their own measurement published in January 2026, roughly three to ten times more tokens than a single-agent approach: the figure attached to this lesson gives that gap in numbers. Their direct recommendation is to start with the simplest approach that works, and to add complexity only when evidence justifies it. A well-equipped single agent often accomplishes far more than you would expect from a lone agent.

This triptych, fan-out, pipeline, barrier, is not the vocabulary Anthropic uses: it is a reading grid specific to this course, built to decide between architectures. The official documentation describes the same moves without naming them this way, it mentions agents that critique each other's work in an adversarial way, or that draft a plan from several angles before comparing them.

Tâche : comparer trois offres fournisseurs et rédiger une recommandation

Fan-out avec barrière :
  agent A analyse l'offre 1
  agent B analyse l'offre 2       vers une synthèse commune
  agent C analyse l'offre 3       qui attend les trois rapports
  synthèse finale (un seul appel, une fois les trois rapports reçus)

Pipeline sans barrière :
  agent A extrait les prix, puis agent B les met en forme, puis agent C les publie
  (chaque agent démarre dès que le précédent a produit son résultat)

Before choosing, ask two questions in order. Can the subtasks genuinely move forward in parallel, without depending on each other. And would a single call, in a single turn, already be enough to produce the expected result. If the answer to the second question is yes, the winning architecture is the one you did not build. The next lesson, on worktree isolation and the dry loop, starts from the assumption that fan-out has already been chosen.