Building Adaptive AI Agents
by DeepLearning.AI
Stop your agent relearning the same lesson every session — build skill induction, code knowledge graphs, and know when to fine-tune.
Overview
Building Adaptive AI Agents is a free one-hour short course from DeepLearning.AI built in partnership with Oracle and taught by Nacho Martinez (Data Scientist Advocate at Oracle) and Casius Lee (AI Developer Advocate at Oracle). It attacks one specific production failure: an agent that solves a task today starts from zero tomorrow, repeating the same mistakes because nothing it learned survives the session. The seven lessons — Introduction, Overview of Continual Learning, Behavior Adaptation: Trajectory-to-Skill Learning, Knowledge Adaptation Through Code Graphs, Building the Code Knowledge Graph, Adapting the Model Itself, and Conclusion — separate three adaptation layers and show when each is the right tool rather than treating memory as one undifferentiated problem. Lesson 3 builds a skill induction pipeline that mines an agent's own traces — conversations, tool calls, errors and the fixes that followed — into reusable skills, with a human in the loop approving each one before it is applied. Lessons 4 and 5 build a code knowledge graph from scratch, linking files through imports, function calls and co-edit history mined from Git, on the stated premise that in a large codebase retrieval, not code generation, is the real bottleneck. Lesson 6 covers weight-space adaptation and when fine-tuning actually pays for itself. Three lessons ship runnable code examples; the reference implementation persists memory and graph structure in Oracle AI Database 26ai. Every video lesson and notebook is free — only the graded quiz and the accomplishment badge require a DeepLearning.AI Pro membership.
At a Glance
- Topic
- Agentic
- Level
- Intermediate
- Format
- Course
- Cost
- Free
- Duration
- ~1 hour 3 minutes of video across 7 lessons, plus 3 code notebooks — about 2-3 hours worked through hands-on
- Provider
- DeepLearning.AI
- Hands-on
- Yes — code/exercises
- Certificate
- None
What You’ll Learn
- ✓Distinguish behavior, knowledge and weight-space adaptation, and pick the right layer per failure
- ✓Build a skill induction pipeline that converts raw agent traces into reusable named skills
- ✓Keep a human in the loop approving every induced skill before an agent may reuse it
- ✓Construct a code knowledge graph from imports, function calls and Git co-edit history
- ✓Query that graph to retrieve context in large repositories instead of regex or naive embedding search
- ✓Judge when fine-tuning the model itself beats adding more context or more skills
- ✓Recognise the four memory forms an adaptive agent needs and where each one is persisted
Highlights
- •Treats continual learning as three separable engineering problems, not one vague 'memory' feature
- •You build the code knowledge graph yourself rather than calling a hosted memory API
- •Explicitly argues retrieval is the bottleneck in large codebases — and then measures it
- •Human-approval step on induced skills is a production safety detail most agent-memory content omits
- •Free, one hour, and taught by Oracle advocates who ship the reference implementation alongside it
Who It’s For
Best For
- ✓AI engineers whose coding agents repeat the same mistakes across sessions
- ✓Teams evaluating agent memory frameworks before committing to one
- ✓Developers working on retrieval quality inside large, unfamiliar codebases
- ✓Engineers deciding whether their next step is more context or a fine-tune
Prerequisites
- •Comfortable writing Python — the hands-on labs assume it
- •Prior experience building at least one tool-calling LLM agent
- •Basic familiarity with embeddings and retrieval concepts
FAQ
What is Building Adaptive AI Agents?
A free short course from DeepLearning.AI and Oracle for engineers whose agents solve a task today and start from zero tomorrow. It separates continual learning into three distinct layers — behavior, knowledge, and weights — and has you build the first two from scratch: a skill induction pipeline that mines agent traces into reusable, human-approved skills, and a code knowledge graph that makes retrieval in a large codebase reflect how the code is actually wired together.
Is Building Adaptive AI Agents free?
Building Adaptive AI Agents is free to access.
What level is Building Adaptive AI Agents for?
Building Adaptive AI Agents is aimed at a intermediate audience. Recommended background: Comfortable writing Python — the hands-on labs assume it, Prior experience building at least one tool-calling LLM agent, Basic familiarity with embeddings and retrieval concepts.
How long does Building Adaptive AI Agents take?
Expect roughly ~1 hour 3 minutes of video across 7 lessons, plus 3 code notebooks — about 2-3 hours worked through hands-on. Most learners work through it at their own pace.
What will I learn from Building Adaptive AI Agents?
You'll learn: Distinguish behavior, knowledge and weight-space adaptation, and pick the right layer per failure; Build a skill induction pipeline that converts raw agent traces into reusable named skills; Keep a human in the loop approving every induced skill before an agent may reuse it; Construct a code knowledge graph from imports, function calls and Git co-edit history; Query that graph to retrieve context in large repositories instead of regex or naive embedding search; Judge when fine-tuning the model itself beats adding more context or more skills; Recognise the four memory forms an adaptive agent needs and where each one is persisted.
Topics
Sources
This page was written from 3 sources, 1 on domains other than deeplearning.ai.