Building Adaptive AI Agents: Skills from Traces and Code Knowledge Graphs

Melvin Vivas · X video post · 2026-08-27 · 2:07 · 109 views · Open on X

Topics: AI Agents, Tool Use & MCP, Retrieval-Augmented Generation (RAG), Prompt & Context Engineering · Level: intermediate

Summary

Melvin Vivas shares the trailer for "Building Adaptive AI Agents," a course built with Oracle and introduced by Andrew Ng. It covers two ways a coding agent can improve between sessions. Behavioral adaptation turns the agent's own traces into reusable skills that a human approves. Knowledge adaptation builds a code knowledge graph that updates as the code changes and finds relevant context that keyword search misses.

Key points

• The problem: agents forget between sessions. A coding agent that spends minutes fixing an environment issue today may repeat the whole process tomorrow, which wastes tokens and time and repeats mistakes.
• Behavioral adaptation: turn the agent's traces (conversations, tool calls, errors and fixes) into reusable 'enhanced skills'.
• A human approves new skills before the agent uses them the next time it meets the same task.
• Knowledge adaptation: build a code knowledge graph that links files through imports, function calls, past changes and co-edits in the repo.
• The agent follows relationships in the graph instead of relying only on keyword search, so it finds the right context more efficiently.
• The graph updates whenever code changes or a new function or file is added, so it is ready for the next retrieval.
• Context Hub, an open-source package from Andrew Ng and collaborators, solves another kind of adaptation problem by letting different agents learn from each other.

Resources mentioned

• Building Adaptive AI Agents · course · deeplearning.ai · free
  A course on agents that learn from experience: turning their traces into reusable skills and building a code knowledge graph for retrieval.
• Context Hub · repo · github.com · free
  An open-source package from Andrew Ng and collaborators that lets different agents learn from each other.
• Andrew Ng · person · x.com · free · recommended by both Bashiri Smith & Melvin Vivas
  A leading AI researcher and educator whose view is that more people should learn to code as AI tools improve.
  Also in: Andrew Ng Says Keep Learning to Code, but Learn the Modern Way (Bashiri Smith on Facebook · notes), Andrew Ng's AI Engineering Skills Map (Melvin Vivas on X · notes), Andrew Ng on how anyone can become an AI Engineer (Melvin Vivas on X · notes)
• Nacho Matinas · person · linkedin.com · free · open in a browser to verify
  Co-instructor of Building Adaptive AI Agents (name as it appears in the machine transcript).
• Cassius Lee · person · blogs.oracle.com · free · open in a browser to verify
  Co-instructor of Building Adaptive AI Agents (name as it appears in the machine transcript).

Try this

• Look up the Building Adaptive AI Agents course and add it to your watchlist.
• Look up the open-source Context Hub package to see how agents can share what they learn.
• Build a pipeline that turns a coding agent's traces (tool calls, errors, fixes) into reusable skills that a human approves before reuse.
• Build a code knowledge graph of a repository (imports, function calls, change history, co-edits) that updates automatically, and use it for agent context retrieval instead of keyword search.