3 sessions on html5 canvas with a coding agent: the agent failed at least once in 3, and 0 finished without a human stepping in.
Takeaway: Next time I'd ask upfront whether a real browser is available in the environment, since the agent-browser testing loop was a dead end here and wasted several turns before falling back to manual code review.
Takeaway: Next time I'd start by telling the agent up front that the repo has a dirty worktree with deleted files, so it doesn't burn a turn getting blocked and asking me to choose a recovery path.
Takeaway: I'd start the first prompt by explicitly telling the agent to run a type-check/build after every file it writes, not just at the end, since most of the wasted turns were chasing TypeScript errors that piled up across files (duplicate TILE_SIZE exports, unused vars, a stray label typo).