Building Adaptive AI Agents: Learning from Experience to Avoid Repeated Mistakes
ORIGINAL / Your coding agent keeps solving the same problem
This course introduces how to build adaptive AI agents that retain experience across sessions, improving performance through skill induction, knowledge graphs, and model fine-tuning, addressing the issue of starting from scratch each time.
01 ABSTRACT
DeepLearning.AI and Oracle present a new course by Nacho Martínez and Casius Lee, covering three adaptation layers for building adaptive AI agents: behavior (skill induction), knowledge (code knowledge graph), and model (fine-tuning). It emphasizes extracting reusable skills from agent traces, building a code knowledge graph for better retrieval, and discussing when fine-tuning is appropriate. The authors argue this is key to building more efficient agents.
02 KEY FINDINGS
- Agents start from zero each session, wasting time and resources.
- Behavior adaptation: build a skill induction pipeline to convert agent traces into reusable skills with human approval.
- Knowledge adaptation: build a code knowledge graph connecting files via imports, function calls, and co-edits in git history to improve context retrieval.
- Model adaptation: learn when to use fine-tuning and other weight-space techniques.
- The course highlights that retrieval, not code generation, is often the bottleneck in large codebases.
AI GENERATED SUMMARY / DISCOVERED BY DEEPLEARNING.AI