LangChain Deep Agents adds subagents, skills

This looks like LangChain making a bet that the winning agent stack is a harness, not just a model wrapper. Deep Agents is basically Claude Code-style ergonomics with a more composable, model-neutral runtime, which is a strong pitch for teams that want control.

• –Subagents keep exploratory work out of the main context window, which matters once an agent starts chaining file reads, searches, and refactors.
• –Skills are progressive-disclosure files, so procedures can be versioned and shared without bloating the base prompt.
• –The filesystem backend makes the system feel restartable and durable, not just conversational, and it can point at local disk or longer-lived stores.
• –The CLI matters because it brings the same stack directly into a terminal workflow, not just into an app embedding.
• –LangChain is positioning Deep Agents as a model-neutral alternative to Claude Agent SDK and Codex-style tooling.