LangChain Unveils Deep Agents Framework for Multi-Agent AI Systems

Zach Anderson
Jan 22, 2026 20:25

LangChain releases Deep Agents with subagents and skills primitives to tackle context bloat in AI systems. Here’s what developers need to know.

LangChain has released Deep Agents, a framework designed to solve one of the thorniest problems in AI agent development: context bloat. The new toolkit introduces two core primitives—subagents and skills—that let developers build multi-agent systems without watching their AI assistants get progressively dumber as context windows fill up.

The timing matters. Enterprise adoption of multi-agent AI is accelerating, with Microsoft publishing new guidance on agent security posture just this week and MuleSoft rolling out Agent Scanners to manage what it calls “enterprise AI chaos.”

The Context Rot Problem

Research from Chroma demonstrates that AI models struggle to complete tasks as their context windows approach capacity—a phenomenon researchers call “context rot.” HumanLayer’s team has a blunter term for it: the “dumb zone.”

Deep Agents attacks this through subagents, which run with isolated context windows. When a main agent needs to perform 20 web searches, it delegates to a subagent that handles the exploratory work internally. The main agent receives only the final summary, not the intermediate noise.

“If the subagent is doing a lot of exploratory work before coming with its final answer, the main agent still only gets the final result, not the 20 tool calls that produced it,” wrote Sydney Runkle and Vivek Trendy in the announcement.

Four Use Cases for Subagents

The framework targets specific pain points developers encounter when building production AI systems:

Context preservation handles multi-step tasks like codebase exploration without cluttering the main agent’s memory. Specialization allows different teams to develop domain-specific subagents with their own instructions and tools. Multi-model flexibility lets developers mix models—perhaps using a smaller, faster model for latency-sensitive subagents. Parallelization runs multiple subagents simultaneously to reduce response times.

The framework includes a built-in “general-purpose” subagent that mirrors the main agent’s capabilities. Developers can use it for context isolation without building specialized behavior from scratch.

Skills: Progressive Disclosure

The second primitive takes a different approach. Instead of loading dozens of tools into an agent’s context upfront, skills let developers define capabilities in SKILL.md files following the agentskills.io specification. The agent sees only skill names and descriptions initially, loading full instructions on demand.

The structure is straightforward: YAML frontmatter for metadata, then a markdown body with detailed instructions. A deployment skill might include test commands, build steps, and verification procedures—but the agent only reads these when it actually needs to deploy.

When to Use What

LangChain’s guidance is practical. Subagents work best for delegating complex multi-step work or providing specialized tools for specific tasks. Skills shine when reusing procedures across agents or managing large tool sets without token bloat.

The patterns aren’t mutually exclusive. Subagents can consume skills to manage their own context windows, and many production systems will likely combine both approaches.

For developers building AI applications, the framework represents a more structured approach to multi-agent architecture. Whether it delivers on the promise of keeping agents out of the “dumb zone” will depend on real-world implementation—but the primitives address problems that anyone building production AI systems has encountered firsthand.

Image source: Shutterstock