5 sessions on gemini api with a coding agent: the agent failed at least once in 2, and 3 finished without a human stepping in.
Takeaway: I'd start by pointing the agent straight at the agents/ folder and the CLI entry points rather than letting it discover the file layout on its own, since that's most of what I actually needed.
Takeaway: I'd start with a more targeted prompt pointing straight at the entry-point scripts (paperbanana.py, app.py, generate_paper_figures.py) instead of asking for a full directory sweep, since most of the value came from the agent files anyway.
Takeaway: I'd start by just asking for a summary of the pipeline architecture and key entry points rather than 'read every script' - it burned a lot of turns reading files sequentially.
Takeaway: Next time I'd tell it upfront to keep the image context small since loading multiple full-resolution figure iterations blew past the model's token limit right at the finish line.
Takeaway: Next time I'd keep the session shorter and more scoped — splitting 'update the paper text' from 'generate all figures and recompile' into separate sessions so the context window doesn't balloon and kill the final compile step.