Somewhere between two alpha tags on July 6th and 7th, LangChain's Deep Agents stopped being a library and became a product line.
The releases page tells the story in package names. There's deepagents itself, now at 0.7.0a6, which the README describes as "the batteries-included agent harness." There's deepagents-code at 0.1.34, which the docs call "a pre-built coding agent in your terminal, similar to Claude Code or Cursor, powered by any LLM." And there's deepagents-acp at 0.0.9, an adapter for the Agent Client Protocol. A fourth, deepagents-talon, rounds out the set.
Three months ago, Deep Agents was one import. Now it's a harness, a shippable coding agent, and a protocol bridge — each versioned on its own clock. That's not a refactor. It's a statement about where the value is.
The two joints nobody used to be able to unstick#
Every coding agent you've used is bundled at two joints.
The first is the model. Claude Code is Claude. Codex is an OpenAI model. The agent's loop — read the repo, plan, edit, run tests, read the errors, edit again — is welded to one provider's weights. The second is the editor. Cursor's agent lives in Cursor. The agent and the surface you drive it from ship as one thing.
Deep Agents pries both joints apart at once. The harness is model-agnostic by design — the README's line is "works with any LLM that supports tool calling: frontier, open-weight, or local," and the 0.7.0a6 alpha went as far as adding a dedicated NVIDIA Nemotron 3 Ultra harness profile alongside the usual OpenAI/Anthropic/Google paths and the self-hosted trio of Ollama, vLLM, and llama.cpp. So the model becomes a config value, not an architecture.
Then deepagents-acp handles the other joint. The Agent Client Protocol is an open, Apache-licensed standard — LSP's idea, applied to coding agents — that standardizes how an editor and an agent talk. Zed shipped it; JetBrains and others have been building to it. Once an agent speaks ACP, it works in any ACP-capable editor without the editor knowing anything about the agent's internals. deepagents-acp is the plug.
The model is a config value. The editor is a client. The only thing Deep Agents keeps for itself is the loop in the middle.
Why the loop is the thing worth keeping#
If you can swap the model and swap the editor, what's left to own? The harness — and that turns out to be the interesting answer.
Read the README's capability list and it's clearly the same harness-shaped core that the whole framework-to-harness migration has been circling: sub-agents that "delegate tasks to agents with isolated context windows," a filesystem tool that reads and writes over "pluggable local, sandboxed, or remote backends," context management that will "summarize long threads and offload tool outputs to disk," plus shell access and persistent cross-session memory. That's the loop-plumbing every serious agent re-implements. Deep Agents' bet is that this plumbing — not the model, not the editor — is the durable, defensible layer.
It's a bet LangChain has been telegraphing for a while, and shipping deepagents-code as a real terminal app makes it concrete. The coding agent isn't a demo notebook anymore; it's a distributable binary with its own release cadence and its own product concerns — the 0.1.33 notes add trust prompts for symlinked skills and selective MCP trust, the kind of thing you only build when people are actually running it against real repos.
The part to keep your head about#
The version numbers are the disclaimer. The harness is at a 0.7.0 alpha — 0.7.0a6, specifically — and the sub-packages sit at 0.1.x and 0.0.x. The deepagents-acp fix in 0.0.9 was "defer interrupt state reads until stream closes," which is exactly the kind of sharp edge you hit when a protocol adapter is young.
So don't pin this in production yet. The reason it matters isn't that it's finished; it's that the shape is now explicit. If the harness owns the loop, the model is swappable, and ACP makes the editor a client, then "which coding agent should we adopt" stops being a single locked decision and becomes three smaller ones you can revisit independently. You can keep your harness and change models when the price moves. You can keep your model and change editors when your team does. The lock-in that made coding-agent choices feel permanent was always in the bundling — and Deep Agents just unbundled it in public, one alpha tag at a time.
