LangChain DeepAgents¶

DeepAgents is LangChain's MIT-licensed agent harness, built on LangGraph. It ships a filesystem backend, an execute tool, subagent delegation, skills, and long-term memory. SuperQode supports it through two separate routes, because DeepAgents ships as two separate products.

Route	Hub entry	Use it when
Deep Agents Code	`deepagents-code`	You want the complete prebuilt terminal coding agent, with its own approvals, memory, and skills
DeepAgents SDK runtime	`deepagents`	You want a SuperQode HarnessSpec whose loop, graph, and subagents are executed by DeepAgents

Both are open source under the MIT license at langchain-ai/deepagents, so both appear under the Harness Hub's Open source filter.

Deep Agents Code¶

deepagents-code is the prebuilt coding agent, published as the dcode command. It exposes itself as an ACP server, which is the route SuperQode connects to.

Install And Authenticate¶

curl -LsSf https://langch.in/dcode | bash

OpenAI, Anthropic, and Gemini providers are included. Select others at install time:

DEEPAGENTS_CODE_EXTRAS="ollama,openrouter" curl -LsSf https://langch.in/dcode | bash

Run dcode once and complete /auth to connect a provider. Then verify that SuperQode can discover it:

superqode agents show deepagents-code
superqode agents doctor deepagents-code

Connect From The TUI¶

:connect deepagents-code

It also appears in the unified Harness Switcher:

:harness
:harness switch deepagents-code
:harness switch deepagents-code --fork

Deep Agents Code owns its agent loop, tools, approval mode, and thread history. SuperQode provides discovery, launch, session continuity, and a normalized event surface.

DeepAgents SDK Runtime¶

The deepagents runtime backend takes a SuperQode HarnessSpec and executes it with create_deep_agent(...). Reach for this when you want DeepAgents graph state, middleware, filesystem backend behavior, and subagent patterns behind a spec you own and can evaluate.

Install¶

uv tool install "superqode[deepagents]"

SuperQode targets the DeepAgents 0.7 API. Any other release reports a version problem rather than a missing package, so an upgrade or downgrade is named directly.

Provider Integration Packages¶

DeepAgents resolves the model with LangChain's init_chat_model, which requires the matching LangChain integration package to be importable. The runtime bundles langchain-anthropic and langchain-google-genai only, so local providers fail with an ImportError such as:

ImportError: Initializing ChatOllama requires the langchain-ollama package.
Please install it with `pip install langchain-ollama`

Install the integration package for your provider into the same tool environment (use --force because superqode is already installed):

uv tool install "superqode[deepagents]" --with langchain-ollama --force

Use the equivalent package for other local providers, for example langchain-openai for LM Studio, MLX, or llama.cpp OpenAI-compatible endpoints. Then re-run :retry in an existing session.

Run The Built-In Preset¶

superqode harness show deepagents
superqode harness run deepagents "Add a regression test for the parser"

Start a repository-owned spec from the same template:

superqode harness init deepagents-coder --template deepagents --output harness.yaml

A fuller example, with an orchestrator workflow and a review subagent, is in examples/harnesses/deepagents.yaml.

What The Spec Controls¶

HarnessSpec field	DeepAgents equivalent
`provider` and `model`	`provider:model` model spec
Job prompt compiled from the spec	`system_prompt`
`agents` after the first, with `delegates_to` or a subagent role	`subagents`
`runtime.config.skills`	`skills`
`runtime.config.memory`	`memory`, defaulting to `context.instruction_files`
`execution_policy.allow_read` and `allow_write`	`FilesystemPermission` deny rules
`working_directory`	`FilesystemBackend(root_dir=...)`

Constraints¶

No-tool harnesses are rejected. The DeepAgents public API is tool-oriented, so use the builtin runtime for model-only specs.
allow_shell: false is rejected. DeepAgents exposes its execute tool whenever a filesystem backend is configured.
DeepAgents runs its own filesystem and execute tools, so SuperQode approvals do not gate them. The Hub states this before you select the harness.

Choose The DeepAgents Route¶

Goal	Route
Use the complete prebuilt coding agent	`:connect deepagents-code`
Run a HarnessSpec you own on the DeepAgents loop	`superqode harness run deepagents "..."`
Run a bare JavaScript Deep Agent over ACP	`:connect acp deepagents`

The third row is the upstream deepagents-acp npm package, which runs a general-purpose Deep Agent rather than the coding agent. It stays in the ACP catalog as a distinct entry.