Deep Agents: LangChain's SDK for Agents That Plan and Delegate

April 1, 2026·1 hr 4 min·5 clips

Sydney Rinkle from LangChain reveals how DeepAgents turn raw LLMs into planning, iterative tools like Cloud Code.

The conversation starts with a plain contrast. Typing into ChatGPT gives the model only the user prompt, while tool-using agents can plan, iterate, test, and recover when the first try misses. That gap is the harness. Sydney Rinkle, back on Talk Python from LangChain, describes DeepAgents as a new open source library for building agents with plain Python functions, middleware hooks, MCP support, and system prompts. Michael keeps grounding the idea. When sub-agents come up, he turns the term into a simple example: one helper reads an article, another reads a document, and both report back. Parallelism matters, but it is not the whole point. The real gain is context isolation, because a small sub-task works better when it gets only the context it needs instead of the whole conversation history. Sydney agrees and ties that back to the system prompt, which tells the agent how to use the filesystem, planning tool, and sub-agents. Memory fits there too. DeepAgents can load memory into the system prompt, giving the harness some continuity across conversations while keeping sub-tasks narrow. Michael catches the practical angle. A sub-agent can turn a messy source into a sentence or two, then pass that back to the main thread without stuffing the context window. The tone is very Talk Python. It gets technical, but Michael keeps checking the listener's mental model with concrete examples instead of letting the vocabulary pile up. The episode is agent architecture, not hype. Planning, delegation, filesystem access, prompt design, and MCP support all sit underneath the familiar chat interface people already know. There is a sponsor wrap at the end. The close points listeners toward Temporal for reliable Python workflows and back to Talk Python courses.

As heard by us

A practical Python-side look at LangChain's Deep Agents and the harness that turns LLMs into planning, tool-using agents.

Deep Agents makes agent building feel less like magic and more like Python infrastructure. Michael Kennedy and Sydney Rinkle focus on the harness around the model: planning tools, file system access, sub-agents, system prompts, memory loaded into those prompts, middleware hooks,…

Read the full review in PlayNext →

Why you'd press play

Want a clearer picture of how agent frameworks work under the hood?

Read the full recommendation in PlayNext →

Listen to the show on

Apple Podcasts Spotify

Play clips from this episode More from Talk Python To Me