deepagents

Overview

Deepagents is an agent framework and, increasingly, a full agent harness for building longer-running AI systems with orchestration logic, tool use, backend integration, and evaluation workflows. Based on recent mentions, its architecture separates the agent loop from a backend that can provide filesystem-like operations and optional sandboxed code execution, and it is built on LangGraph. It also supports standard endpoints like MCP and a2a, shared interfaces such as web and Slack, and even “fake” backends for non-coding agents.

For AI Product Managers, deepagents matters because it reflects a broader shift from lightweight framework abstractions toward operational harnesses designed for real-world agent behavior: context selection, tool availability, failure recovery, subagent coordination, and measurement. Recent updates—especially dynamic subagents and dynamically loading tools when a skill is loaded—suggest a modular approach that can improve maintainability while preserving prompt cache efficiency with OpenAI or Anthropic models.

Key Developments

• 2026-04-30: Harrison Chase said he was optimizing deepagents for peak performance on OSS models, motivated by a view that closed-model costs may become prohibitively high.
• 2026-06-01: He shared how to evaluate DeepAgents at scale on AWS with LangSmith, including datapoint and evaluator design for longer-horizon agents.
• 2026-06-21: Deepagents was highlighted as part of a nearly 10-hour agentic AI course covering LangChain, LangGraph, RAG, and guardrails.
• 2026-06-30: Harrison Chase introduced dynamic subagents in deepagents, enabling developers to programmatically spin up subagents across multiple use cases.
• 2026-07-01: He demonstrated a live voice agent pattern that pairs Gemini Live for low-latency interaction with DeepAgents for more complex reasoning.
• 2026-07-06: Chase described an industry shift from frameworks like LangChain, AI SDK, and LlamaIndex toward fuller harnesses such as DeepAgents, Claude Agent SDK, and EVE.
• 2026-07-10: He clarified that DeepAgents is not runtime lock-in, describing it as OS-based and portable across environments including SuperQode, LangGraph, Temporal, and others.
• 2026-08-17: Chase recapped the architecture: agent loop separated from a backend with filesystem-like operations and optional sandboxed code execution; support for local or cloud deployment, MCP and a2a endpoints, and shared web and Slack interfaces.
• 2026-09-10: He said a loop is being built around deepagents for orchestration logic and LangSmith for measurement, emphasizing that agent improvement depends heavily on harness quality at the tool boundary.
• 2026-10-08: Deepagents added support for dynamically loading tools when a skill is loaded, allowing skill-specific tools to be paired without always exposing them and, with OpenAI or Anthropic models, potentially preserving prompt cache efficiency.

Relevance to AI PMs

• Design more modular agent products: The new pattern of loading tools only when a skill is loaded points to a cleaner way to structure capabilities. PMs can use this to reduce unnecessary tool exposure, simplify permissioning, and keep product features organized by skill or workflow.
• Plan for evaluation and operations early: Deepagents is repeatedly discussed alongside LangSmith and large-scale evaluation. For PMs, this reinforces that agent product quality depends not just on prompts or model choice, but on measurement design, failure recovery, and orchestration behavior.
• Choose architectures that stay portable: Mentions of OSS model optimization, runtime portability, MCP/a2a compatibility, and local or cloud deployment suggest a strategy for avoiding lock-in. PMs can use this as a lens when prioritizing infrastructure bets and vendor dependencies.

Related

• Harrison Chase: The main public voice associated with deepagents, sharing product direction, architecture, and usage patterns.
• LangGraph: Identified as the underlying foundation deepagents is built on.
• LangSmith: Frequently paired with deepagents for evaluation, measurement, and improvement loops.
• Claude Agent SDK / claude-agent-sdk: Mentioned as a comparable harness in the shift from frameworks to fuller agent systems.
• LangChain, AI SDK, LlamaIndex: Referenced as earlier-style frameworks contrasted with deepagents as a fuller harness.
• MCP and a2a: Standard endpoints that deepagents supports, relevant for interoperability.
• Anthropic, Claude, and OpenAI: Model providers mentioned in connection with prompt cache behavior and broader agent ecosystem comparisons.
• AWS: Referenced in the context of evaluating DeepAgents at scale with LangSmith.
• Guardrails: Mentioned alongside deepagents in educational material, reflecting its role in production agent design.
• Gemini Live: Used in an example where conversational latency is handled separately from deeper reasoning delegated to DeepAgents.
• OSS models: A stated optimization target for deepagents, tied to cost and performance strategy.
• Temporal, SuperQode, EVE: Related through portability and comparisons with alternative runtimes or agent harness approaches.
• memory and mcpsskillssubagents: Adjacent themes connected to longer-horizon agents, modular skills, and subagent patterns.