LangChain releases Deep Agents
LangChain released 'Deep Agents', a structured runtime for multi‑step AI agents with planning, memory and context isolation — turning ephemeral prompts into long‑lived, stateful workflows reported. That changes system design constraints: expect questions about agent persistence, context sharding and safe scaling of long‑running AI tasks.
LangChain announced Deep Agents on March 15, 2026 as an open‑source “agent harness” for long‑running workflows, with the project repository published on GitHub. (marktechpost.com) The runtime is explicitly built on LangChain primitives and the LangGraph execution layer for durable, streaming execution and human‑in‑the‑loop workflows rather than a new model backend. (docs.langchain.com) Key built‑in primitives listed in the repo include a planner (write_todos) for task decomposition, a filesystem with read_file/write_file/edit_file/ls/glob/grep, the ability to spawn subagents via a task API, and a sandboxed execute tool for shell commands. (github.com) By default the agent filesystem is transient and scoped to a single thread, while the docs show a CompositeBackend pattern to route specific paths to persistent stores for long‑term memory. (docs.langchain.com) LangChain added an autonomous context‑compression capability (announced March 11, 2026) that lets agents trigger compression rules instead of waiting for fixed token thresholds, addressing mid‑workflow context loss. (blockchain.news) There are both Python and JS/TS packages (deepagents and deepagentsjs), and example projects (including a deep_research demo) demonstrate running via a LangGraph server, Jupyter notebooks, and a CLI that expects API keys for models and search services. (github.com) LangChain positions Deep Agents as a production‑oriented harness that forces explicit decisions about context sharding, hybrid persistence (route /memory paths to durable backends), and orchestration complexity—an approach the CEO described as “harness engineering” over model‑only solutions. (docs.langchain.com)
