3 sessions on scryfall api with a coding agent: the agent failed at least once in 2, and 1 finished without a human stepping in.
Takeaway: I'd start the next session by pasting in the actual SOW questions up front instead of doing three rounds of back-and-forth, since most of the ambiguity was around color computation rules and navigation UX that I could have specified from the start.
Takeaway: Next time I'd point the agent straight at .github/workflows/run-crawler.yml and entrypoint.py instead of letting it search broadly, since .cursorignore blocked the workflow file and cost a lot of exploration turns.
Takeaway: Next time I'd front-load the dataset ID and scope questions in the first prompt so the agent doesn't have to burn a dozen turns guessing at what I meant by 'seed_dict' vs full retrain.