LangChain has published a detailed framework for building multi-agent AI systems, arriving as the AI infrastructure space heats up with competing approaches from Google and Microsoft in recent weeks.

The guide, authored by Sydney Runkle, identifies four core architectural patterns that developers can use when single-agent systems hit their limits. The timing isn’t accidental—Google released its own eight essential multi-agent design patterns on January 5, while Microsoft unveiled its Agentic Framework just days ago on January 14.

When Single Agents Break Down

LangChain’s position is clear: don’t rush into multi-agent architectures. Start with a single agent and good prompt engineering. But two constraints eventually force the transition.

Context management becomes the first bottleneck. Specialized knowledge for multiple capabilities simply won’t fit in a single prompt. The second constraint is organizational—different teams need to develop and maintain capabilities independently, and monolithic agent prompts become unmanageable across team boundaries.

Anthropic’s research validates the approach. Their multi-agent system using Claude Opus 4 as lead agent with Claude Sonnet 4 subagents outperformed single-agent Claude Opus 4 by 90.2% on internal research evaluations. The key advantage: parallel reasoning across separate context windows.

The Four Patterns

Subagents use centralized orchestration. A supervisor agent calls specialized subagents as tools, maintaining conversation context while subagents remain stateless. Best for personal assistants coordinating calendar, email, and CRM operations. The tradeoff: one extra model call per interaction.

Skills take a lighter approach—progressive disclosure for agent capabilities. The agent loads specialized prompts and knowledge on-demand rather than managing multiple agent instances. LangChain controversially calls this a “quasi-multi-agent architecture.” Works well for coding agents where context accumulates but capabilities stay fluid.

Handoffs enable state-driven transitions where the active agent changes based on conversation context. Customer support flows that collect information in stages fit this pattern. More stateful, requiring careful management, but enables natural multi-turn conversations.

Routers classify input and dispatch to specialized agents in parallel, synthesizing results. Enterprise knowledge bases querying multiple sources simultaneously benefit here. Stateless by design, which means consistent per-request performance but repeated routing overhead for conversations.

Performance Numbers That Matter

LangChain’s benchmarks reveal concrete tradeoffs. For a simple one-shot request like “buy coffee,” Handoffs, Skills, and Router each require 3 model calls. Subagents needs 4—that extra call provides centralized control.

Repeat requests show where statefulness pays off. Skills and Handoffs save 40% of calls on the second identical request by maintaining context. Subagents maintains consistent cost through stateless design.

Multi-domain queries expose the biggest divergence. Comparing Python, JavaScript, and Rust documentation (2000 tokens each), Subagents processes around 9K total tokens while Skills balloons to 15K due to context accumulation—a 67% difference.

What Developers Should Consider

The framework arrives as multi-agent systems move from research curiosity to production requirement. LangChain’s Deep Agents offers an out-of-the-box implementation combining subagents and skills for teams wanting to start quickly.

But the core advice remains pragmatic: add tools before adding agents. Graduate to multi-agent patterns only when you hit clear limits. The 90% performance gains Anthropic demonstrated are real, but so is the complexity overhead of coordinating multiple AI agents in production environments.

Image source: Shutterstock