DeepAgents
A ready-to-use Agent Harness built on top of LangChain and LangGraph, providing production-grade modules including a planner, tool-call retry mechanism, and context compression. The framework is designed for multi-step, long-chain complex agent tasks, enabling developers to deploy capable agents without building core infrastructure from scratch.
Core Facts
DeepSeek Harness offers multiple working modes, forming a spectrum between autonomy and controllability.
Use LangChain for simple chained tasks, LangGraph for complex branching logic, and choose DeepAgents when enterprise-grade capabilities such as file system access, sandboxing, and skill invocation are needed.
Agent development frameworks include LangChain, LangGraph, and DeepAgents within the LangChain ecosystem, as well as Spring AI and Spring AI Alibaba for Java developers.
Timeline (last 90 days)
The long-term stable operation capability emphasized by DeepAgents is largely built on LangGraph's state management and persistence capabilities.
DeepAgents is still undergoing rapid iteration, and sub-agents are its core abstraction.
Learning DeepAgents requires focusing on filling in LangGraph's state flow and branch control mechanisms, as well as the creation and scheduling patterns of sub-agents.
The LangChain team launched the DeepAgents framework around November 2025.
When designing asynchronous sub-agents, four types of concurrency issues need to be considered: result aggregation order, error propagation strategy, state consistency, and resources and rate limiting.
DeepAgents supports two modes of creating synchronous sub-agents and asynchronous sub-agents: in synchronous mode the main agent waits for the sub-agent to finish before continuing; in asynchronous mode the sub-agents can execute in parallel, and the main agent does not have to block and wait.
Even without a solid foundation in LangChain/LangGraph, users are not hindered from understanding and using DeepAgents, because the underlying process mechanisms have been abstracted and encapsulated.
DeepAgents primarily addresses coordination and lifecycle management across multiple intelligent agents, freeing developers from manually writing graph node and edge transition logic.
According to the presenter, enterprise-level projects are currently adopting DeepAgents' multi-agent collaboration architecture for implementation.
In long-cycle tasks, asynchronous mode can significantly reduce overall execution time, while synchronous mode ensures result correctness when logic is strictly sequential; the two modes can be used in combination.
26 more timeline events
All Facts (20)
DeepSeek Harness offers multiple working modes, forming a spectrum between autonomy and controllability.
80%VerifiedUse LangChain for simple chained tasks, LangGraph for complex branching logic, and choose DeepAgents when enterprise-grade capabilities such as file system access, sandboxing, and skill invocation are needed.
75%VerifiedAgent development frameworks include LangChain, LangGraph, and DeepAgents within the LangChain ecosystem, as well as Spring AI and Spring AI Alibaba for Java developers.
75%VerifiedThe early DeepAgents framework is a complete Harness architecture implementation that built approximately seven engineering modules around its core: tool invocation, file system, sandbox environment, context management, memory system, logic orchestration and middleware, and feedback loops with constraint mechanisms.
70%VerifiedThe LangChain ecosystem has a three-layer structure: LangGraph at the bottom layer, LangChain in the middle layer, and DeepAgents at the top layer.
70%VerifiedWhen a synchronous sub-agent is called by the main Agent, it blocks the main process and waits for the result before continuing, making it suitable for scenarios where results have dependencies on one another.
65%VerifiedBoth DeepAgents and LangChain are implemented based on LangGraph.
65%VerifiedThe sub-agent pattern breaks complex tasks into multiple independent subtasks, where each sub-agent has an isolated context and returns results to the orchestrator upon completion.
65%VerifiedDeepAgents is a wrapper built on top of LangChain and LangGraph.
65%Verified智能体（Agent）是具备自主决策能力的AI，能自行拆解任务如搜索信息、提取内容、生成报告
65%UnverifiedDeepAgents的图结构设计支持有向图工作流、持久化状态检查点和Human-in-the-loop人机协作模式
85%UnverifiedDeepAgents项目采用Python开发，可直接对接Hugging Face、OpenAI SDK等主流工具库
80%UnverifiedAutoGPT、CrewAI、MetaGPT等Agent框架更侧重于提供预定义的Agent角色和协作模式，而DeepAgents更强调底层运行时的健壮性和可控性
70%UnverifiedDeepAgents is a relatively new framework in the LangChain ecosystem, positioned toward enterprise-level deployment, and supports capabilities such as local file system operations, sandbox mechanisms, and Skill invocation.
60%UnverifiedIn LangChain's hosted Deep Agents, the model, prompts, tools, and sub-agents are determined by the developer, while persistence, memory, skill loading, code execution sandboxing, and deployment are handled by the platform.
60%UnverifiedThe core of DeepAgents is the main entry function create_deep_agent, which is used to create agents with planning, tool-calling, and context management capabilities.
60%UnverifiedDeepAgents is still undergoing rapid iteration, and sub-agents are its core abstraction.
50%UnverifiedDeepAgents supports two modes of creating synchronous sub-agents and asynchronous sub-agents: in synchronous mode the main agent waits for the sub-agent to finish before continuing; in asynchronous mode the sub-agents can execute in parallel, and the main agent does not have to block and wait.
50%UnverifiedWhen designing asynchronous sub-agents, four types of concurrency issues need to be considered: result aggregation order, error propagation strategy, state consistency, and resources and rate limiting.
50%UnverifiedThe long-term stable operation capability emphasized by DeepAgents is largely built on LangGraph's state management and persistence capabilities.
50%