Deep Agents (2026): LangChain Harness, Planning, Subagents & Deep Research

1. Deep Agents from LangChain

9 entries

The open-source agent harness and its documentation, code and companion tools. Start with the overview, then the quickstart and repository.

01Deep Agents overview (LangChain docs)Official documentation for the Deep Agents harness: planning, filesystem, subagents, skills and memory. 02Deep Agents quickstartBuild and run a first deep agent. 03Deep Agents on GitHubSource repository for the Python package and command-line agent. 04deepagents on PyPIThe Python package, installed with pip install deepagents. 05Deep Agents for JavaScriptTypeScript and JavaScript version of the harness. 06LangChain Blog: Deep AgentsThe original post explaining why agents need planning, files, subagents and detailed prompts. 07LangChain Blog: Agent Frameworks, Runtimes and HarnessesExplains where frameworks, runtimes and harnesses differ. 08LangChain Blog: Doubling Down on DeepAgentsLangChain's follow-up on the project's direction. 09LangChain AcademyFree courses on LangChain, LangGraph and agents.

2. Where Deep Agents Fit: LangChain, LangGraph and LangSmith

10 entries

The layers underneath and around the harness. LangChain supplies primitives, LangGraph the runtime and Deep Agents the opinionated harness on top.

10LangChain documentationPrimitives and integrations for models, tools and agents. 11LangGraph documentationLow-level runtime for stateful, graph-based agent workflows. 12LangGraph on GitHubSource for the LangGraph runtime. 13LangChain on GitHubSource for LangChain. 14LangGraph PlatformDeploy and scale long-running agents. 15LangSmithTracing, evaluation and monitoring for agents. 16LangGraph: PersistenceCheckpoints that let long tasks pause, resume and recover. 17LangGraph: Human-in-the-LoopPause an agent for approval before sensitive actions. 18Model Context ProtocolOpen standard for connecting agents to tools and data. 19LangChain MCP AdaptersUse MCP tools inside LangChain and LangGraph agents.

3. Core Ideas: Planning, Filesystems, Subagents, Skills and Memory

12 entries

The design principles that make an agent 'deep', with sources. Each idea addresses a failure of simple tool-calling loops on long tasks.

20Anthropic: Effective Context Engineering for AI AgentsCompaction, note-taking and sub-agents for long-horizon work. 21Anthropic: How We Built Our Multi-Agent Research SystemOrchestrator and subagent design with lessons on prompts and cost. 22Anthropic: Equipping Agents with Agent SkillsProgressive disclosure of skills loaded on demand. 23Agent Skills specificationOpen format for packaging reusable agent skills. 24Manus: Context Engineering for AI AgentsPractical lessons including using the file system as context. 25Cognition: Don't Build Multi-AgentsA caution on sharing context between subagents. 26ReAct: Synergizing Reasoning and ActingThe reason-and-act loop behind most tool-calling agents. 27Plan-and-Solve PromptingPlanning before executing to reduce missing steps. 28ReflexionAgents learn from feedback stored in memory. 29Voyager: An Open-Ended Embodied AgentA skill library that an agent grows over time. 30MemGPT: Towards LLMs as Operating SystemsPaging memory between context and storage. 31Lost in the MiddleWhy long contexts degrade and why offloading helps.

4. Deep Research Agents

8 entries

Products and projects that apply the deep-agent pattern to research. Long-running research is the clearest use case for planning and subagents.

32OpenAI Deep ResearchAgent that browses and analyses sources to produce cited reports. 33Gemini Deep ResearchGoogle's multi-step research agent. 34LangChain Open Deep ResearchOpen-source configurable deep research agent. 35GPT ResearcherOpen-source autonomous research agent. 36Perplexity Deep ResearchPerplexity's report-generating research mode. 37Deep Research Agents: A Systematic Examination and RoadmapSurvey of architecture, tools and benchmarks for deep research agents. 38Humanity's Last ExamHard benchmark often used to test research agents. 39BrowseCompBenchmark for browsing agents that find hard-to-locate facts.

5. Agent Harnesses, SDKs and Coding Agents

13 entries

Other harnesses and SDKs that package long-running agent loops. Compare how each handles planning, context, subagents and sandboxes.

40Claude Agent SDKAnthropic's SDK for building agents on the Claude Code harness. 41Claude CodeTerminal coding agent that popularised todo lists, subagents and skills. 42OpenAI Agents SDKLightweight framework with handoffs, guardrails and tracing. 43Google Agent Development KitCode-first toolkit for multi-agent systems. 44Microsoft Agent FrameworkSuccessor to AutoGen and Semantic Kernel agent features. 45OpenHandsOpen-source platform for software development agents. 46smolagentsSmall library for agents that write code to act. 47Pydantic AITyped Python agent framework. 48CrewAIFramework for role-based multi-agent teams. 49MastraTypeScript framework for agents, workflows and memory. 50LettaPlatform for stateful agents with persistent memory. 51GooseOpen-source on-machine agent from Block. 52ClineOpen-source coding agent for the editor.

6. Memory, Sandboxes and Runtime Tooling

9 entries

Components a deep agent relies on: long-term memory, safe code execution and storage. Long tasks need somewhere safe to run and something to remember with.

53LangMemLong-term memory tools for LangGraph agents. 54Mem0Memory layer for AI agents. 55Zep and GraphitiTemporal knowledge graph memory for agents. 56E2BCloud sandboxes for running agent-generated code. 57DaytonaSecure sandbox infrastructure for AI code. 58Modal SandboxesRun untrusted code in isolated containers. 59TavilySearch API built for agents. 60ExaSearch engine API for AI applications. 61FirecrawlTurns websites into clean data for agents.

7. Related Pages on This Site

10 entries

Neighbouring directories for building, evaluating and securing agents. Deep agents combine patterns covered across these pages.

62Top Agentic AI FrameworksFrameworks for building agents. 63AI Agent FrameworksAgent development frameworks compared. 64Multi-Agent System PatternsDesign, orchestration and decomposition patterns. 65Multi-Agent Orchestration FrameworksFrameworks for coordinating several agents. 66Evaluating & Improving Agentic ApplicationsEvaluation frameworks, benchmarks and optimisation for agents. 67AI Agent SkillsReusable skills for agents. 68AI Agent ObservabilityMetrics, events, logs and traces for agents. 69AI Agent GuardrailsFive checkpoints and tools. 70Advanced Architectural Patterns for Agentic AIArchitecture patterns for agentic systems. 71Agentic AIAgentic AI hub.

Related Harness References

1 entries

Further references.

72Terminal-BenchBenchmark.

0 entries

Related research directories on this site.

Frequently Asked Questions

What are deep agents?

Deep agents are AI agents built to work autonomously on complex, multi-step tasks that run for hours, such as deep code refactoring, comprehensive research and multi-system workflows. Unlike single-turn chatbots or simple prompt chains, they add explicit planning, a filesystem for context, subagents, skills and persistent memory.

What is LangChain Deep Agents?

Deep Agents is an open-source agent harness from LangChain. It is an opinionated, batteries-included framework that packages planning, a virtual filesystem, subagent delegation, skills and context management so developers do not have to build them from scratch. It is installed with pip install deepagents.

What are the core features of a deep agent?

Five layers recur: explicit planning and task decomposition with todo lists, a virtual filesystem for offloading large outputs, subagents with isolated context, reusable skills loaded on demand, and persistent long-term memory across sessions.

How do Deep Agents relate to LangChain and LangGraph?

LangChain provides the primitives and model and tool integrations. LangGraph provides the low-level runtime for stateful, graph-based workflows. Deep Agents sit on top as an agent harness that packages planning, filesystems, subagents and context engineering.

Why do deep agents use a virtual filesystem?

Stuffing every document, code file and tool output into the context window exhausts it and degrades results. A filesystem lets the agent write large content to disk and read back only what it needs, while middleware compresses history and supports prompt caching.

What are subagents and why use them?

A subagent is an isolated child agent with its own scoped context that handles one chunk of work, often in parallel, and reports back to the parent. This keeps the main agent's memory uncluttered. Sharing too little context between subagents can cause conflicting decisions, so design handoffs carefully.

What are deep agents used for?

Typical uses are deep research that produces cited reports, large code changes, data and document analysis, and workflows that span several systems. Short, simple tasks usually do not need a deep agent and run cheaper with a plain tool-calling loop.

Which additional references does this page add?

The section 'Related Harness References' lists Claude Agent SDK, Terminal-Bench, Agent Skills.

How current is this list?

Links were checked in November 2026. A few sites, such as OpenAI's, block automated checks and may need a manual click. This page is a resource list, not engineering advice, and tool descriptions reflect each project's own positioning.