When to use multi-agent systems (and when not to)

Author: Cara Phillips, with Paul Chen, Andy Schumeister, Brad Abrams, Theo Chu (Anthropic) · Type: practitioner · Status: draft · Published: 2026-01-23 · URL: https://claude.com/blog/building-multi-agent-systems-when-and-how-to-use-them

Licence: Copyright Anthropic; free to read, quote with attribution.

Summary

Anthropic guidance on when multi-agent architectures are justified: context protection, parallelization, and specialization. Warns that elaborate planner/executor/reviewer pipelines often lose context at handoffs and burn 3-10x more tokens than a single well-tooled agent; recommends starting with the simplest design that works.

Insights

• Default to a single capable agent with focused tools; add multi-agent structure only when a task genuinely decomposes into parallel independent threads. (high, draft) orchestration cost agents in production

  Start with the simplest approach that works, and add complexity only when evidence supports it. — Recommendations / conclusion

• Elaborate multi-stage agent pipelines frequently underperform because context is lost at each handoff and coordination costs exceed execution costs. (high, draft) orchestration cost

  We’ve observed teams build elaborate multi-agent systems with separate agents for planning, execution, review, and iteration, only to discover that they suffered from lost context at each handoff and spent more tokens coordinating than executing. — When not to use multi-agent systems

• Tool sprawl degrades agent performance; keeping each agent’s toolset small is a cheap quality lever. (medium, draft) orchestration agents in production

  When an agent has access to too many tools (often 20+), performance suffers. — Design guidance on tools