Fable Brain

Multi-Agent Fan-Out Research Flow

4 min read

Multi-Agent Fan-Out Research Flow

Fan out parallel worker agents from a lead orchestrator only for breadth-first research worth roughly 15x chat token spend; Anthropic’s own orchestrator-worker Research system “outperformed single-agent Claude Opus 4 by 90.2%” on its internal eval (https://www.anthropic.com/engineering/multi-agent-research-system).

What it is

The orchestrator-workers pattern, one of five composable workflow patterns cataloged in Anthropic’s Building Effective Agents essay (December 2024) alongside prompt chaining, routing, parallelization, and evaluator-optimizer. Anthropic’s Research feature realizes it: a lead Opus 4 agent plans and spawns parallel Sonnet 4 subagents, each in its own context window. Anthropic distinguishes workflows, where LLMs and tools follow predefined code paths, from agents, where the LLM dynamically directs its own process; this flow sits at the boundary, with a dynamic orchestrator over parallel workers.

How it works

  1. Qualify the task. Multi-agent fits breadth-first, parallelizable research that exceeds a single context window. It fits poorly when subagents need shared context or tight real-time coordination, which is why most coding tasks stayed single-agent in 2025.
  2. Decompose. The lead agent plans slices and writes each worker a bounded brief: scope, sources, expected output shape.
  3. Fan out. Spawn workers in parallel, each in a separate context window so research does not consume the orchestrator’s context.
  4. Collect summaries. Workers return condensed findings; the orchestrator keeps requirements, decisions, and integration.
  5. Synthesize and verify. The lead merges results and runs verification before presenting, ideally in fresh context.

In Claude Code the mechanics are native: subagents are Markdown files with YAML frontmatter (name, description, tools, model) stored in .claude/agents/ or ~/.claude/agents/, and “Each subagent runs in its own context window with a custom system prompt” (https://code.claude.com/docs/en/sub-agents). Built-in Explore, Plan, and general-purpose subagents are delegated to automatically, so light fan-out happens even without custom workers.

Key insight

Token spend is the dominant variable: in Anthropic’s analysis it alone explained 80 percent of performance variance. Fan-out buys performance by buying tokens, which is exactly why the qualification step exists.

Best practice

  • Reserve multi-agent for high-value tasks; systems “use about 15× more tokens than chats” (single agents roughly 4x), and token spend alone explained 80 percent of performance variance in Anthropic’s analysis. EVIDENCE-BASED
  • Start with the simplest thing: direct LLM API calls first, agentic complexity only when it demonstrably improves outcomes. “Maintain simplicity in your agent’s design” (https://www.anthropic.com/engineering/building-effective-agents). EVIDENCE-BASED
  • Keep tightly coupled work single-agent; shared-context and real-time coordination needs break the fan-out model. EVIDENCE-BASED
  • Prototype fan-out batch runs on 2 to 3 files or slices before scaling to the full set. PRACTITIONER
  • Have workers return high-signal summaries rather than raw dumps, mirroring the tool-response guidance to return token-efficient, high-signal content. PRACTITIONER
  • For autonomous agent teams (the February 2026 C compiler ran 16 parallel Opus 4.6 agents coordinating through git and file locks), treat near-perfect task verifiers as the critical enabler. CONTESTED

Pitfalls

  • Paying 15x tokens for a task a single agent handles; qualify before fanning out.
  • Splitting a tightly coupled task; workers cannot see each other’s context.
  • Framework abstractions that obscure prompts and hinder debugging, warned against in Building Effective Agents.
  • Unbounded worker briefs; without scope and output shape, synthesis quality collapses.
  • Trusting the C compiler result as routine practice; it is a single vendor experiment, hence CONTESTED here.
  • Fanning out without a synthesis owner; the orchestrator, not the human, must hold requirements and integration.

Sources

Next actions

  • Write a reusable worker-brief template: scope, allowed sources, output schema, summary length cap.
  • Measure one real research task single-agent versus fan-out and record token cost against quality.
  • Define the qualification checklist (breadth-first? exceeds one window? high value?) as a pre-flight gate.

Graph View

Backlinks