Notes · September 13, 2026
I tried the course's code graph on my own backend
I finished the DeepLearning.AI course Building Adaptive AI Agents. I really liked it, and then did what usually makes a course more useful for me: I tried one of its ideas on my own backend.
The course separates adaptation into three layers:
Behavior: turn an agent’s previous traces into reusable skills.
Knowledge: build a code graph so the agent can find relevant code through imports, function calls, and Git co-edits.
The model itself: adapt its weights when the context is already right, but the agent still does not behave as expected.
The second one interested me most. The promise is practical: help the coding agent reach the right files sooner, take fewer steps, and use fewer tokens.
I implemented the course approach for the My Storyland backend: MiniLM finds semantic starting points, then PageRank follows the code connections. I tested Codex with and without that map on two historical bugs.
On the simple bug, both agents produced the correct fix, but the graph version used 5.5% more input tokens.
On the harder security bug, both agents passed all tests. The graph version, though, took about half a minute longer and used almost twice as many input tokens.
Why?
MiniLM found the correct authentication function. PageRank then gave too much weight to highly connected files and didn’t put the three routers that needed changes in the first 12 results. Codex had to find them itself anyway.
My backend is fairly small, so Codex can find the relevant files quickly without a graph. There is simply less room for the graph to help.
I still think the idea can save time and tokens in a larger repository, but the graph needs its own evaluation.
My next version would reduce the influence of Git co-edits and combine semantic similarity with PageRank instead of letting PageRank completely reorder the results.
I have not tried the third layer yet: adapting the model itself with LoRA.
My Storyland marketing agent may be a good place to test it. Sometimes it has the right source material and detailed instructions, but it still does something different from what I expect—and adding one more line to the prompt doesn’t change that.
Course: Building Adaptive AI Agents Certificate: deeplearning.ai