GenAI PMdeepagents
An agent framework/architecture that separates the agent loop from backend operations like filesystem access and optional sandboxed code execution. Useful for building agents with flexible local/cloud deployment and multiple interfaces.
Key Highlights
- Deepagents separates the agent loop from backend operations like filesystem access and optional sandboxed code execution.
- Its architecture supports local or cloud deployment, standard interfaces like MCP and A2A, and multiple product surfaces such as web, Slack, and voice.
- The project has been positioned as a portable agent harness rather than a runtime-locked platform.
- Deepagents has expanded from an open-source Claude Agent SDK into a broader pattern for evaluation, subagents, and OSS model optimization.
deepagents
Overview
Deepagents is an agent framework and architecture for building AI agents where the core agent loop is separated from backend operations such as filesystem-style access and optional sandboxed code execution. In practice, this means teams can design an agent’s reasoning and orchestration independently from the runtime environment it uses to read files, call tools, or execute code. It has been described as open source, built on LangGraph, and capable of running across local or cloud deployments with support for standard endpoints like MCP and A2A.For AI Product Managers, deepagents matters because it represents a shift from lightweight prompting frameworks toward fuller agent harnesses that are more production-oriented. Its architecture suggests better portability, more flexible deployment choices, and clearer separation of concerns between agent intelligence and operational infrastructure. That makes it relevant when evaluating how to ship durable agent products across web, Slack, voice, internal tooling, or enterprise environments without tightly coupling product behavior to a single runtime stack.
Key Developments
- 2026-04-10: Harrison Chase noted that community middleware such as langchain-task-steering was emerging to customize agents and deepagents, signaling early ecosystem experimentation around extensibility.
- 2026-04-13: Deepagents launched as an open-source Claude Agent SDK, positioned for building and running Claude-powered agents.
- 2026-04-30: Harrison Chase said he was optimizing deepagents for strong performance on open-source models, reflecting interest in cost control and model portability.
- 2026-06-01: Chase outlined how to evaluate DeepAgents at scale on AWS with LangSmith, including datapoint and evaluator design for long-horizon agents.
- 2026-06-21: Deepagents was highlighted as part of a long-form agentic AI course covering LangChain, LangGraph, RAG, and guardrails, indicating growing educational adoption.
- 2026-06-30: Dynamic subagents were introduced in Deepagents, allowing developers to programmatically create subagents for multiple use cases.
- 2026-07-01: Chase demonstrated a live voice agent pattern using Gemini Live for low-latency interaction while offloading more complex reasoning to DeepAgents.
- 2026-07-06: Chase framed the market as moving from frameworks like LangChain, AI SDK, and LlamaIndex toward more complete agent harnesses such as DeepAgents, Claude Agent SDK, and EVE.
- 2026-07-10: Chase clarified that DeepAgents is not runtime lock-in and can run across different environments including SuperQode, LangGraph, Temporal, or other platforms.
- 2026-08-17: Chase recapped deepagents’ architecture: a separated agent loop plus backend, optional sandboxed execution, support for local and cloud setups, MCP and A2A endpoints, shared web and Slack interfaces, and “fake” backends for non-coding agents.
Relevance to AI PMs
1. Evaluate architecture choices early. Deepagents provides a concrete pattern for separating agent reasoning from operational backends. PMs can use this to ask better questions about portability, security boundaries, code execution needs, and deployment flexibility before committing to a stack.2. Plan multi-surface agent products. Because deepagents has been discussed in the context of web, Slack, local/cloud, and voice workflows, it is useful for PMs designing one agent experience that needs to work across multiple interfaces without rebuilding core logic.
3. Improve cost, testing, and model strategy. Mentions of OSS model optimization and LangSmith-based evaluation suggest a practical PM playbook: reduce inference costs, benchmark long-horizon behavior, and design evaluations that reflect real user workflows rather than simple one-turn accuracy.
Related
- Harrison Chase: The main public voice behind deepagents, sharing architecture, launch updates, and usage patterns.
- LangGraph: Deepagents is described as being built on LangGraph, connecting it to graph-based orchestration for agents.
- Claude / Anthropic / Claude Agent SDK: Deepagents launched as an open-source Claude Agent SDK for Claude-powered agents.
- LangSmith / langsmith-agent-builder: Relevant for evaluation, observability, and agent development workflows around deepagents.
- langchain-task-steering: An example of community middleware used to customize agent behavior.
- memory / mcpsskillssubagents: Related to agent capabilities and patterns, especially as deepagents expands into dynamic subagents and more modular systems.
- guardrails: Frequently relevant in the same agentic stack for safety, control, and reliability.
- OSS models: A notable optimization focus for deepagents, especially for cost-sensitive deployments.
- AWS: Mentioned in the context of evaluating deepagents at scale.
- Gemini Live: Used alongside DeepAgents in a live voice agent pattern where interaction and reasoning are separated.
- AI SDK, LlamaIndex, EVE: Comparative reference points in the shift from traditional frameworks to fuller agent harnesses.
- SuperQode, Temporal: Cited as examples showing DeepAgents’ runtime portability.
- MCP and A2A: Standard endpoints/interfaces that help position deepagents within a broader interoperable agent ecosystem.
Newsletter Mentions (11)
“Harrison Chase recapped deepagents’ architecture, which separates the agent loop from a backend providing filesystem-like operations and, optionally, sandboxed code execution. Built on LangGraph, it supports local or cloud setups, standard endpoints such as MCP and a2a, shared web and Slack interfaces, and “fake” backends for non-coding agents.”
#1 𝕏 Harrison Chase recapped deepagents’ architecture, which separates the agent loop from a backend providing filesystem-like operations and, optionally, sandboxed code execution. Built on LangGraph, it supports local or cloud setups, standard endpoints such as MCP and a2a, shared web and Slack interfaces, and “fake” backends for non-coding agents.
“Harrison Chase clarifies that DeepAgents isn’t runtime lock-in—being OS-based, you can run it anywhere, whether in SuperQode with a different runtime, LangGraph, Temporal, or any other platform.”
The item frames DeepAgents as a portable agent system rather than a platform-locked one.
“Harrison Chase observes the agent industry pivoting from frameworks like LangChain, AI SDK, and LlamaIndex to full-fledged harnesses such as DeepAgents, Claude Agent SDK, and EVE—with DeepAgents predating EVE by about ten months.”
#2 𝕏 Harrison Chase observes the agent industry pivoting from frameworks like LangChain, AI SDK, and LlamaIndex to full-fledged harnesses such as DeepAgents, Claude Agent SDK, and EVE—with DeepAgents predating EVE by about ten months.
“Harrison Chase shows how to build a live voice agent by offloading complex reasoning to DeepAgents and using Gemini Live for natural, low-latency interactions.”
Harrison Chase shows how to build a live voice agent by offloading complex reasoning to DeepAgents and using Gemini Live for natural, low-latency interactions. #15 📝 Claude Code Blog Getting started with loops - A tutorial-style post introducing loops in Claude Code, aimed at helping developers get started using loop constructs and workflows.
“#8 𝕏 Harrison Chase introduced dynamic subagents in Deepagents, letting you programmatically spin up subagents and showcasing six distinct use cases.”
#8 𝕏 Harrison Chase introduced dynamic subagents in Deepagents, letting you programmatically spin up subagents and showcasing six distinct use cases.
“Harrison Chase highlights a nearly 10-hour agentic AI course covering LangChain, LangGraph, RAG, deepagents and guardrails.”
#3 𝕏 Harrison Chase highlights a nearly 10-hour agentic AI course covering LangChain, LangGraph, RAG, deepagents and guardrails. He’s also asking for other strong Lang* resources for learners.
“Harrison Chase breaks down how to evaluate DeepAgents at scale on AWS with LangSmith, covering concrete datapoint and evaluator design methods for longer-horizon agents.”
#2 𝕏 Harrison Chase breaks down how to evaluate DeepAgents at scale on AWS with LangSmith, covering concrete datapoint and evaluator design methods for longer-horizon agents.
“#19 𝕏 Harrison Chase predicts that by 2026 closed-model costs will be prohibitively high and he’s optimizing deepagents for peak performance on OSS models.”
#19 𝕏 Harrison Chase predicts that by 2026 closed-model costs will be prohibitively high and he’s optimizing deepagents for peak performance on OSS models. #20 in Peter Yang showcases five new Google Labs AI products—Pomelli for marketing, Stitch for design, Genie for 3D worlds, Flow for video creation, and NotebookLM for research synthesis.
“#1 𝕏 Harrison Chase launched Deepagents—an open-source Claude Agent SDK available at github.com/langchain-ai/deepagents for building and running Claude-powered agents.”
GenAI PM Daily April 13, 2026 GenAI PM Daily 🎧 Listen to this brief 3 min listen Today's top 14 insights for PM Builders, ranked by relevance from X, Blogs, and YouTube. Deepagents Releases Open Source Claude Agent SDK #1 𝕏 Harrison Chase launched Deepagents—an open-source Claude Agent SDK available at github.com/langchain-ai/deepagents for building and running Claude-powered agents.
“Harrison Chase notes that community middleware—like “langchain-task-steering”—is popping up for customizing agents and deepagents, and invites anyone with middleware to contribute by reaching out to Sydney.”
#25 𝕏 Harrison Chase notes that community middleware—like “langchain-task-steering”—is popping up for customizing agents and deepagents, and invites anyone with middleware to contribute by reaching out to Sydney.
Related
An AI company best known for Claude. It is referenced implicitly through Claude’s memory and Cowork features.
Anthropic’s assistant, discussed here for shared memory across chat and Cowork. The feature is relevant to PMs because it enables cross-task context reuse and user-controlled memory.
An AI infrastructure company and community that recapped a founder dinner in San Francisco. The discussion focused on vertical agents, moats, and go-to-market implications.
Founder and builder in the AI agents ecosystem, associated here with explaining deepagents’ architecture and its use of LangGraph. Relevant to PMs for understanding agent-loop/backend separation and deployment patterns.
An interoperability protocol for connecting AI systems and tools. Here it is described through a public roadmap covering long-running workloads, local-server HTTP, discovery, identities, permissions, and generated SDKs.
A framework for building LLM applications and agents. In this newsletter it appears in the story about the founders’ attempt to automate dropshipping.
A developer/evaluation tool cited in benchmark testing of automated eval systems. The newsletter uses it as part of a comparison against harder-to-detect product-judgment failures.
An SDK for building Claude-based agents and workflows. It is cited as one of the newer harness-style tools replacing older frameworks.
Amazon’s cloud platform, referenced in a story about a training pipeline running on a 4-GPU instance. The anecdote highlights GPU utilization monitoring and infrastructure waste.
Vercel’s SDK for integrating AI features into apps. The newsletter highlights token savings from a single line of code in DeepSeek-powered workflows.
A standard endpoint/protocol for agent-to-agent or agent interoperability, mentioned here alongside MCP as a supported interface. It matters to PMs as part of agent connectivity and integration strategy.
Community middleware example for customizing agent behavior and steering tasks in agent frameworks.
Stay updated on deepagents
Get curated AI PM insights delivered daily — covering this and 1,000+ other sources.
Subscribe Free