DeepAgents
Source: https://docs.langchain.com/oss/python/deepagents/overview (fetched 2026-07-06 — re-verify if this content seems stale; this is a fast-moving 2026-era product).
What It Is, and Isn't
DeepAgents is not a separate orchestration engine — it's an opinionated harness layer built on top of LangChain's create_agent and LangGraph's runtime. Think of the stack as three layers: LangGraph (low-level execution runtime — state, checkpointing), LangChain's create_agent (compose exactly the primitives you need), DeepAgents' create_deep_agent (a pre-assembled, batteries-included harness on top of both). Reach for DeepAgents when a task needs the harness's defaults (filesystem, planning, subagents, context management); reach for plain create_agent (langchain-code) when you want to compose a narrower agent by hand instead.
Core Architecture — Four Pillars
- Execution Environment: custom tools/APIs/MCP servers, a virtual filesystem (pluggable backends — in-memory, local disk, LangGraph store, composite routing), declarative glob-based filesystem permission rules, code execution via sandboxes.
- Context Management:
SKILL.md-based progressive-disclosure skills, AGENTS.md persistent memory loaded at startup, automatic summarization/context offloading for long-running work, automatic prompt caching for Anthropic/Bedrock models.
- Delegation:
write_todos for structured task tracking, a task tool for spawning ephemeral, stateless subagents with fresh isolated context that return one final report each (not a back-and-forth dialogue).
- Steering: human-in-the-loop via LangGraph interrupts, configurable per tool.
Key API
from deepagents import create_deep_agent
agent = create_deep_agent(
model="anthropic:claude-sonnet-4-6",
tools=[custom_functions],
system_prompt="...",
memory=[memory_files], # AGENTS.md-style persistent memory
permissions=[permission_rules], # glob-based, first-match-wins filesystem access control
interrupt_on={"edit_file": True},
)
agent.invoke({"messages": [{"role": "user", "content": "..."}]})
See references/architecture-and-api.md for the default middleware stack, HarnessProfile, and every gotcha below in full detail.
Gotchas (do not skip)
- You cannot remove the default middleware — only hide specific tools from a model via
HarnessProfile's excluded_tools. Removing the middleware itself is "intentionally rejected" by the framework's own design.
- Subagents are stateless — a spawned subagent returns exactly one final report, not a streaming dialogue. Don't design a flow expecting back-and-forth with a subagent mid-task.
- Filesystem permission rules don't cover sandbox backends — a sandbox allows arbitrary shell execution regardless of glob-based permission rules on the built-in filesystem tools. Add custom policy hooks if a sandbox needs additional guardrails.
- Prompt caching is automatic for Anthropic/Bedrock — no configuration needed; don't add manual cache-control logic on top of it.
1---2name: deepagents-code3description: Build long-running, production-ready agents with DeepAgents — the opinionated harness on top of LangGraph/LangChain providing filesystem access, task planning, subagent delegation, and context management out of the box. Use for long-running coding/research/multi-step tasks where you'd otherwise hand-assemble that middleware stack yourself; use langchain-code's plain create_agent instead for a simple single-tool-call loop with no need for planning or filesystem access. Triggers on "DeepAgents", "create_deep_agent", "agent harness", "long-running agent", "coding agent with filesystem", "subagent spawning".4---56# DeepAgents78Source: https://docs.langchain.com/oss/python/deepagents/overview (fetched 2026-07-06 — re-verify if this content seems stale; this is a fast-moving 2026-era product).910## What It Is, and Isn't1112DeepAgents is **not** a separate orchestration engine — it's an opinionated harness layer built on top of LangChain's `create_agent` and LangGraph's runtime. Think of the stack as three layers: LangGraph (low-level execution runtime — state, checkpointing), LangChain's `create_agent` (compose exactly the primitives you need), DeepAgents' `create_deep_agent` (a pre-assembled, batteries-included harness on top of both). Reach for DeepAgents when a task needs the harness's defaults (filesystem, planning, subagents, context management); reach for plain `create_agent` (`langchain-code`) when you want to compose a narrower agent by hand instead.1314## Core Architecture — Four Pillars15161. **Execution Environment**: custom tools/APIs/MCP servers, a virtual filesystem (pluggable backends — in-memory, local disk, LangGraph store, composite routing), declarative glob-based filesystem permission rules, code execution via sandboxes.172. **Context Management**: `SKILL.md`-based progressive-disclosure skills, `AGENTS.md` persistent memory loaded at startup, automatic summarization/context offloading for long-running work, automatic prompt caching for Anthropic/Bedrock models.183. **Delegation**: `write_todos` for structured task tracking, a `task` tool for spawning ephemeral, stateless subagents with fresh isolated context that return one final report each (not a back-and-forth dialogue).194. **Steering**: human-in-the-loop via LangGraph interrupts, configurable per tool.2021## Key API2223```python24from deepagents import create_deep_agent2526agent = create_deep_agent(27 model="anthropic:claude-sonnet-4-6",28 tools=[custom_functions],29 system_prompt="...",30 memory=[memory_files], # AGENTS.md-style persistent memory31 permissions=[permission_rules], # glob-based, first-match-wins filesystem access control32 interrupt_on={"edit_file": True},33)3435agent.invoke({"messages": [{"role": "user", "content": "..."}]})36```3738See `references/architecture-and-api.md` for the default middleware stack, `HarnessProfile`, and every gotcha below in full detail.3940## Gotchas (do not skip)4142- **You cannot remove the default middleware** — only hide specific tools from a model via `HarnessProfile`'s `excluded_tools`. Removing the middleware itself is "intentionally rejected" by the framework's own design.43- **Subagents are stateless** — a spawned subagent returns exactly one final report, not a streaming dialogue. Don't design a flow expecting back-and-forth with a subagent mid-task.44- **Filesystem permission rules don't cover sandbox backends** — a sandbox allows arbitrary shell execution regardless of glob-based permission rules on the built-in filesystem tools. Add custom policy hooks if a sandbox needs additional guardrails.45- **Prompt caching is automatic for Anthropic/Bedrock** — no configuration needed; don't add manual cache-control logic on top of it.
cwijayasundara