deepagents

Overview

DeepAgents is an open-source agent harness and orchestration layer for building longer-running agent loops, control logic, and multi-step execution flows. It was introduced by Harrison Chase as a Claude Agent SDK, but newsletter coverage increasingly frames it more broadly as a portable system for running agents across environments rather than as a narrowly model- or runtime-bound framework. In practice, DeepAgents sits in the part of the stack that manages how agents act over time: selecting tools, coordinating subagents, handling retries or failure recovery, and connecting to execution backends.

For AI Product Managers, DeepAgents matters because it represents the shift from simple prompt chains and developer frameworks toward full agent harnesses. The newsletter positions it alongside LangSmith for measurement, with DeepAgents handling orchestration logic and LangSmith handling evaluation and observability. That separation is important for PMs designing production agent products: the product challenge is no longer just model quality, but how the system structures tasks, manages memory and tools, supports different runtimes, and improves through feedback loops.

Key Developments

• 2026-04-13: Harrison Chase launched DeepAgents as an open-source Claude Agent SDK for building and running Claude-powered agents.
• 2026-04-30: Chase said he was optimizing DeepAgents for strong performance on open-source models, anticipating rising costs for closed models.
• 2026-06-01: He outlined 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: Dynamic subagents were introduced in DeepAgents, enabling developers to programmatically spin up subagents for multiple use cases.
• 2026-07-01: Chase showed a live voice agent architecture that used Gemini Live for low-latency interaction while offloading complex reasoning to DeepAgents.
• 2026-07-06: He 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: Chase clarified that DeepAgents does not imply runtime lock-in and can run across environments such as SuperQode, LangGraph, Temporal, or other platforms.
• 2026-08-17: He recapped DeepAgents’ architecture: an agent loop separated from a backend that provides filesystem-like operations and optional sandboxed code execution, with support for local or cloud deployment, MCP and a2a endpoints, shared web and Slack interfaces, and fake backends for non-coding agents.
• 2026-09-10: Chase said a broader loop is being built around DeepAgents for orchestration logic and LangSmith for measurement, reinforcing the view that agent improvement depends on harness quality.

Relevance to AI PMs

• Designing production agent systems: DeepAgents is useful as a reference for how modern agent products are structured beyond prompting—around loops, tool use, backends, subagents, and recovery logic. PMs can use this framing to define product requirements at the harness level, not just the model level.
• Evaluating long-horizon behavior: The pairing of DeepAgents with LangSmith shows a practical split between orchestration and measurement. PMs responsible for quality can use this model to separate execution logic from evaluation pipelines, making it easier to test task completion, tool usage, and failure modes at scale.
• Planning for portability and cost: Newsletter mentions emphasize support for OSS models, multiple runtimes, and standard interfaces like MCP and a2a. For PMs, that makes DeepAgents relevant when reducing vendor lock-in, controlling inference cost, or designing products that may need to switch models and execution environments over time.

Related

• Harrison Chase: Primary advocate and source of most newsletter mentions; positioned DeepAgents as part of the shift toward agent harnesses.
• LangSmith / langsmith-agent-builder: Frequently paired with DeepAgents as the measurement and evaluation layer for agent systems.
• LangGraph: DeepAgents is described as being built on LangGraph, tying it to graph-based workflow and stateful execution patterns.
• Anthropic / Claude / claude-agent-sdk: DeepAgents launched as an open-source Claude Agent SDK and remains closely associated with Claude-powered agent development.
• LangChain, AI SDK, LlamaIndex: These are contrasted with DeepAgents as earlier-generation frameworks versus fuller agent harnesses.
• Guardrails, memory, mcpsskillssubagents: These relate to the operational concerns DeepAgents helps manage, including safety, persistence, and modular agent capabilities.
• AWS: Mentioned in the context of evaluating DeepAgents at scale with LangSmith.
• Gemini Live: Used in an example architecture where low-latency voice interaction is separated from heavier reasoning handled by DeepAgents.
• Temporal, SuperQode, EVE: Referenced to emphasize portability and to compare DeepAgents with other runtimes or harness-style systems.
• MCP and a2a: Standard endpoints DeepAgents reportedly supports, relevant for interoperability with external tools and agent ecosystems.