deepagents

deepagents.profiles exposes beta APIs that may receive minor changes in future releases. Refer to the versioning documentation for more details.

Registrations are additive: if a profile is already registered under key (including a built-in profile loaded during lazy bootstrap), the new profile is merged on top rather than replacing it. The incoming profile's fields win on conflicts; unspecified fields inherit from the existing profile. pre_init callables chain (existing runs first), and init_kwargs_factory callables chain — both factories are invoked at every resolution (base first, then override) and their outputs merge with the override's values winning on shared keys.

To layer additional kwargs onto a built-in profile, register under the same provider key. To override a built-in default (e.g. disable the OpenAI Responses API), set the conflicting key explicitly:

from deepagents import ProviderProfile, register_provider_profile

# Adds temperature alongside the built-in `use_responses_api=True`.
register_provider_profile("openai", ProviderProfile(init_kwargs={"temperature": 0}))

# Explicitly disables Responses API for OpenAI. (This will break usage,
# this example is purely illustrative.)
register_provider_profile(
    "openai",
    ProviderProfile(init_kwargs={"use_responses_api": False}),
)

Classes

DeepAgentState

AgentState with DeltaChannel on messages to reduce checkpoint growth from O(N²) to O(N).

AsyncSubAgent

Specification for an async subagent running on a remote Agent Protocol server.

Async subagents connect to any Agent Protocol-compliant server via the LangGraph SDK. They run as background tasks that the main agent can monitor and update.

Compatible with LangGraph Platform / LangSmith Deployment (managed) and self-hosted servers.

Authentication for LangGraph Platform / LangSmith Deployment is handled automatically by the SDK via environment variables (LANGGRAPH_API_KEY, LANGSMITH_API_KEY, or LANGCHAIN_API_KEY). For self-hosted servers, pass custom auth via headers.

AsyncSubAgentMiddleware

Middleware for async subagents running on remote Agent Protocol servers.

This middleware adds tools for launching, monitoring, and updating background tasks on remote Agent Protocol servers. Unlike the synchronous SubAgentMiddleware, async subagents return immediately with a task ID, allowing the main agent to continue working while subagents execute.

Works with any Agent Protocol-compliant server — LangGraph Platform (managed) or self-hosted (e.g. a FastAPI server implementing the Agent Protocol spec).

Task IDs are persisted in the agent state under async_tasks so they survive context compaction/offloading and can be accessed programmatically.

Files can be stored using any backend that implements the BackendProtocol.

FilesystemMiddleware

Middleware for providing filesystem and optional execution tools to an agent.

This middleware adds filesystem tools to the agent: ls, read_file, write_file, edit_file, glob, and grep.

If the backend implements SandboxBackendProtocol, an execute tool is also added for running shell commands.

This middleware also automatically evicts large tool results to the file system when they exceed a token threshold, preventing context window saturation.

FilesystemPermission

A single access rule for filesystem operations.

MemoryMiddleware

Middleware for loading agent memory from AGENTS.md files.

Loads memory content from configured sources and injects into the system prompt. Supports multiple sources that are combined together. See constructor for the full argument list.

Observing non-satisfied terminations

RubricMiddleware

Middleware that drives self-evaluated iteration against a rubric.

The middleware activates only when a caller passes a rubric on invocation state. With no rubric, both before_agent and after_agent return without modifying state, so the middleware is safe to include unconditionally in a create_deep_agent stack.

When grading ends with failed, max_iterations_reached, or grader_error, the middleware does not mutate the response messages. The last AIMessage in the agent's output is whatever the model produced just before the grader gave up. Callers who need to branch on non-satisfied termination must inspect one of:

_rubric_status on the returned state (or agent.get_state(...) on a checkpointed thread),
the on_evaluation callback,
the rubric_evaluation_end stream event.

An info log is also emitted when max_iterations_reached fires.

CompiledSubAgent

A pre-compiled agent spec.

Note

The runnable's state schema must include a 'messages' key.

This is required for the subagent to communicate results back to the main agent.

Note

CompiledSubAgent runnables are used as provided. They do not inherit create_deep_agent(state_schema=...); if the runnable needs custom state fields, compile it with a compatible state schema yourself.

When the subagent completes, the parent reads the returned state: if structured_response is non-None, it is JSON-serialized and used as the ToolMessage content; otherwise, the last non-empty AIMessage text is used.

SubAgent

Specification for an agent.

When using create_deep_agent, subagents automatically receive a default middleware stack before any custom middleware specified in this spec.

SubAgentMiddleware

Middleware for providing subagents to an agent via a task tool.

This middleware adds a task tool to the agent that can be used to invoke subagents.

Subagents are useful for handling complex tasks that require multiple steps, or tasks that require a lot of context to resolve.

A chief benefit of subagents is that they can handle multi-step tasks, and then return a clean, concise response to the main agent.

Subagents are also great for different domains of expertise that require a narrower subset of tools and focus.

GeneralPurposeSubagentProfile

Edits applied to the auto-added general-purpose subagent.

deepagents.profiles exposes beta APIs that may receive minor changes in future releases. Refer to the versioning documentation for more details.

These settings only affect the default subagent that create_deep_agent inserts when the caller does not explicitly provide a subagent named general-purpose.

HarnessProfile

Runtime configuration for deep agent behavior.

deepagents.profiles exposes beta APIs that may receive minor changes in future releases. Refer to the versioning documentation for more details.

A HarnessProfile describes prompt-assembly, tool visibility, middleware, and default-subagent adjustments applied by create_deep_agent once a chat model has been constructed. Profiles are registered via register_harness_profile under a provider key ("openai") or a full provider:model key ("openai:gpt-5.4").

This complements ProviderProfile, which controls the model-construction phase (e.g. init_chat_model kwargs, pre-init side effects). Concerns that shape how the model is built belong in ProviderProfile; concerns that shape how the agent runs belong here.

For YAML/JSON-backed profiles, use HarnessProfileConfig, which contains only the declarative subset and can be passed directly to register_harness_profile.

The extra_middleware field expects langchain.agents.middleware.types.AgentMiddleware instances or a factory returning a sequence of them.

HarnessProfileConfig

Declarative harness-profile config for YAML/JSON-backed profiles.

deepagents.profiles exposes beta APIs that may receive minor changes in future releases. Refer to the versioning documentation for more details.

A HarnessProfileConfig contains the file-friendly subset of harness settings: plain strings, bools, lists, and nested dicts that can be loaded from YAML or JSON. For in-code/runtime-only adjustments such as extra_middleware or class-form excluded_middleware, use HarnessProfile instead.

Class-path serialization

excluded_middleware in config files currently only accepts plain middleware-name strings matched against AgentMiddleware.name. A future revision may add explicit class-path (module:Class) entries for excluding middleware whose class isn't part of the public import surface; until then, exclude such middleware via its .name (using serialized_name for stable public aliases) or stay on the runtime HarnessProfile and pass the class directly.

Config objects may be passed directly to register_harness_profile; the helper converts them to runtime HarnessProfile objects automatically.

ProviderProfile

Declarative configuration for constructing a chat model.

deepagents.profiles exposes beta APIs that may receive minor changes in future releases. Refer to the versioning documentation for more details.

A ProviderProfile describes provider- or model-specific kwargs, pre-initialization side effects, and runtime-derived kwargs that should be applied when resolve_model turns a string spec (e.g. "openai:gpt-5.4") into a BaseChatModel. Profiles are registered via register_provider_profile under a provider key ("openai") or a full provider:model key ("openai:gpt-5.4").

Profiles handle model-construction concerns only — things that shape how init_chat_model assembles the client instance. Typical examples: constructor kwargs like use_responses_api, temperature, max_tokens, or base_url; provider-specific headers such as OpenRouter app attribution; pre-construction checks like minimum-version enforcement; and env-var-aware defaults.

Runtime and harness behavior — system-prompt assembly, tool description overrides, excluded tools, extra middleware, general-purpose subagent configuration — belongs in HarnessProfile, the separate harness profile system consumed by create_deep_agent, not here.

Modules

graph

Primary graph assembly module for Deep Agents.

Provides create_deep_agent, the main entry point for constructing a fully configured deep agent with planning, filesystem, subagent, and summarization middleware.

profiles

Public beta APIs for model and harness profiles.

deepagents.profiles exposes beta APIs that may receive minor changes in future releases. Refer to the versioning documentation for more details.

Profiles let Deep Agents tailor behavior to a specific provider or model spec across two orthogonal phases:

Provider profiles (ProviderProfile) control the model-construction phase. They declare how resolve_model builds a chat model — init_chat_model kwargs, pre-initialization side effects, and kwargs derived from runtime state (e.g. environment variables).
Harness profiles (HarnessProfile, HarnessProfileConfig) control the runtime phase. They declare how create_deep_agent shapes the agent after the model is built — prompt assembly, tool visibility, middleware, and default subagent behavior.

Both kinds live in keyed registries that accept a provider or provider:model key. Registration helpers (register_provider_profile, register_harness_profile) are additive: re-registering under an existing key merges on top of the prior registration rather than replacing it.

Directory layout:

provider/ — ProviderProfile API (provider_profiles.py) plus built-in provider modules (e.g. _openai, _openrouter).
harness/ — HarnessProfile API (harness_profiles.py) plus built-in harness modules for frontier model specs (e.g. _anthropic_sonnet_4_6, _openai_codex).
_builtin_profiles.py — bootstrap that registers built-in profiles and loads third-party plugins (via importlib.metadata entry points) lazily on first profile-registry access, so importing this package stays cheap.
_keys.py — shared validation and lookup helpers for the provider / provider:model registry keys used by both registries.

middleware

Middleware for the Deep Agents agent.

Overview

The LLM receives tools through two paths:

SDK middleware (this package) -- tools, system-prompt injection, and request interception that any SDK consumer gets automatically.
Consumer-provided tools -- plain callable functions passed via the tools parameter to create_deep_agent(). The CLI uses this path for lightweight, consumer-specific tools.

Both are merged by create_deep_agent() into the final tool set the LLM sees.

Why middleware instead of plain tools?

Middleware subclasses AgentMiddleware, overriding its wrap_model_call() hook that intercepts every LLM request before it is sent. This lets middleware:

Filter tools dynamically -- e.g. FilesystemMiddleware removes the execute tool at call-time when the resolved backend doesn't support it.
Inject system-prompt context -- e.g. MemoryMiddleware and SkillsMiddleware inject relevant instructions into the system message on every call so the LLM knows how to use the tools they provide.
Transform messages -- e.g. SummarizationMiddleware counts tokens, truncates old tool arguments, and replaces history with summaries when the context window fills up.
Maintain cross-turn state -- middleware can read/write a typed state dict that persists across agent turns (e.g. summarization events).

A plain tool function in a tools=[] list cannot do any of this -- it is only invoked by the LLM, not before the LLM call.

When to use each path

Use middleware when the tool needs to:

Modify the system prompt or tool list per-call
Track state across turns
Be available to all SDK consumers (not just the CLI)

Use a plain tool when:

The function is stateless and self-contained
No system-prompt or request modification is needed
The tool is specific to a single consumer (e.g. CLI-only)