Fan-out and fan-in (parallelization)
Splitting work into parallel model calls or subagents (fan-out) and merging their results in one step (fan-in).
Fan-out sends parts of a task to several model calls or subagents at the same time. Fan-in collects their outputs and merges them into one result. In an agentic workflow this is the parallelization pattern, and it comes in two forms: sectioning and voting.
Sectioning splits a task into independent subtasks, such as one reviewer per file or one researcher per source, so the work finishes sooner and each call has a small, focused context. Voting runs the same task several times and combines the answers, which adds confidence to a judgment like “is this code vulnerable?” Both forms are described in Anthropic’s Building effective agents. Claude Code dynamic workflows expose them as script functions: parallel() runs a set of agent tasks at once and pipeline() runs one agent per item in a list (docs). Cost grows with the number of branches, because each branch pays for its own tokens.
The weak point is fan-in. Results overlap, contradict each other or come back empty, and a merge step that only concatenates them hides all of that. A good fan-in deduplicates, ranks, resolves conflicts and reports which branches failed. When the split itself has to be decided at run time, the pattern becomes orchestrator-workers. When you keep the best of several attempts instead of merging them, see best-of-n.
Sources
- Anthropic: Building effective AI agents anthropic.com
- Claude Code Docs: Orchestrate subagents at scale with dynamic workflows code.claude.com
Related terms
- Agentic workflow (workflow vs agent): A process where your code fixes the steps and a model does some of them, as opposed to an agent that decides its own steps.
- Multi-agent system (multi-agent orchestration): Several AI agents with separate roles and contexts working on one task, usually coordinated by a lead agent or a fixed workflow.
- Orchestrator-workers pattern (supervisor): One model plans the task, hands parts of it to worker models or subagents, and combines their results.
- Subagent: A separate agent run with its own fresh context, started by a main agent for one subtask.
- Best-of-N sampling: Generating several answers to the same prompt and keeping the best one, chosen by a test, a score or another model acting as judge.
- Dynamic workflow: An orchestration script the agent writes at run time. A runtime executes it, starting many subagents in phases and combining their results in code.
- Prompt chaining: A fixed sequence of model calls where each step works on the previous step’s output, with checks in code between the steps.
- Tokens: The pieces of text models read and write, and the unit you pay for.
- Agent swarm: A mostly marketing name for many AI agents working in parallel on one task; in practice usually a coordinated multi-agent system.
