Learning Skills with Deepagents

The biggest gap between AI agents and human intelligence is the ability to learn. There are various emerging approaches to support continual learning for AI agents. Here, we discuss skill-learning using the DeepAgents skill creator. We reflect over Deepagent trajectories and use this to learn skills. 

Learn how to build Deep Agents on LangChain Academy:
https://academy.langchain.com/courses/deep-agents-with-langgraph/?utm_medium=social&utm_source=youtube&utm_campaign=q4-2025_youtube-academy-links_aw

Skill creator skill:
https://github.com/langchain-ai/deepagents/pull/579

LagnSmith fetch:
https://github.com/langchain-ai/langsmith-fetch

Video notes:
https://www.notion.so/Continual-Learning-with-DeepAgents-2ce808527b1780c79205c176d7b6887a?source=copy_link

Chapters:
0:00 - Introduction: Continual learning and the gap between humans and AI agents
0:27 - Learning approaches: Weight updates vs. learning in context
0:54 - The continual learning loop: Trajectories, reflection, and three update types
1:30 - Three reflection flavors: Memories, instructions, and skills
1:54 - Related research: JPA for prompt optimization, memory updating, skill learning
2:44 - Setting up the Skill Creator skill from the Deep Agents repo
3:24 - Configuring environment variables for dual tracing projects
3:54 - Fetching recent threads with LangSmith Fetch
4:34 - Demo: Starting Deep Agents with skill creator loaded
5:04 - Demo: Reflecting on a thread to create a new skill automatically
6:18 - Reviewing the generated skill: YAML front matter and validation
7:00 - Testing the new skill: Verifying it loads in future sessions
7:24 - Recap: Skill learning as the third leg of continual learning