LangChain Deep Agents (March 15 2026): write_todos Decomposition + task-Isolated Subagents + Filesystem Context Offload

MarkTechPost·high signal

LangChain shipped Deep Agents on March 15—an agent harness built on LangGraph with three built-in primitives: `write_todos` for upfront multi-step task decomposition, a `task` tool that spawns subagents with isolated context windows, and filesystem tools that externalize large context to disk rather than accumulating it in the prompt window. This directly addresses the verification bottleneck by separating planning (main agent) from execution (isolated subagents) while keeping the main context clean. It is the production-grade alternative to naive single-agent loops for tasks exceeding 10+ steps.