Deep Agents

Use this skill to help build Deep Agents applications that stay understandable as they gain planning, files, memory, and delegation.

Deep Agents is not the default answer for every agent. Start by checking whether the task actually needs the harness: multi-step planning, large context, filesystem-backed work, persistent memory, or isolated subagents. If not, prefer a simpler agent or a direct LangGraph workflow.

Working Rules

1. Use the current official docs as source of truth. Before exploring Deep Agents documentation, fetch the LangChain docs index at https://docs.langchain.com/llms.txt, then open only the relevant Python pages for the task. Read official-docs-map.md when choosing pages.
2. Treat local course material as examples, not API truth. Use it for mental models, pedagogy, and runnable examples after checking the current docs. Read local-course-map.md when the user wants examples or teaching support.
3. Keep the architecture honest. Decide explicitly on:
   • why this needs Deep Agents instead of a lighter alternative
   • what belongs in tools vs memory vs skills
   • which backend stores which files
   • whether subagents, permissions, human approval, or code execution are actually needed
4. Build the smallest runnable harness first. Start with a working agent, then add one capability at a time: tools, backend, permissions, memory/skills, subagents, human-in-the-loop, code execution.
5. Prefer context isolation over prompt inflation. Use files, references, and subagents to keep the main agent focused instead of stuffing every instruction and intermediate result into the top-level prompt.

Decision Flow