Talk Python To Me #543: Deep Agents: LangChain's SDK for Agents That Plan and Delegate
82 snips
Apr 1, 2026 Sydney Runkle, a LangChain developer who builds open source agent tooling, walks through Deep Agents and how they make models plan, iterate, and recover. She discusses planning and to-do systems, file system context, sub-agents and parallel work, middleware hooks, MCP interoperability, and using any model or local runtimes. Practical demos, CLIs, and safety sandboxes round out the conversation.
AI Snips
Chapters
Transcript
Episode notes
Social Media Agent Helps Edit Not Replace Posts
- Sydney experimented with a social media deep agent that studies past posts and suggests new ones tuned to performance.
- She prefers writing original posts and using the agent for tough edits, showing agents as style assistants not full authors.
Automatic Tool Schema From Python Signatures
- Deep Agents auto-generates tool schemas from Python signatures, type hints, and docstrings so the model knows which args to pass.
- Sydney notes they convert signatures to JSON schema (using Pydantic) and include that in the prompt to the model.
MCP Lets Agents Use External Tool Catalogs
- Deep Agents supports MCP so agents can discover and call tools hosted by external MCP servers, expanding available capabilities.
- Sydney explains fetching tools from MCP servers allows cross-team or community-shared tool usage inside agents.