6 sessions on conda-smithy with a coding agent: the agent failed at least once in 3, and 3 finished without a human stepping in.
Takeaway: Next time I'd start by dumping the installed packages in the build prefix earlier, since the duplicate systemlib packages were the actual root cause and cost a lot of back-and-forth with conda/python/pixi path issues.
Takeaway: Next time I'd tell the agent upfront to inspect the actual extracted source tree before writing a patch, since guessing at line numbers from a stale copy wasted two build cycles.
Takeaway: I'd tell the agent upfront to check conda-smithy's GITHUB_ACTIONS_RUNS_ON mapping directly instead of spending several web searches first, since the source code fetch is what actually gave the answer.
Takeaway: Next time I'd tell it up front that macOS ships BSD install, not GNU coreutils, since that's the kind of platform detail it can't discover just by reading the recipe skill and docs.
Takeaway: I'd start the prompt by pointing straight at the conda-forge skill's go-licenses reference since that's exactly where the fix came from — it saved a lot of guesswork.
Takeaway: Next time I'd start by telling the agent straight away that the failure is likely a go-licenses classification issue for a Go dependency, since that's a known conda-forge pattern.