GoTourThree takes on the same Go tutorial.

By the numbers ​

Wall-clock between the agent's first phase commit and its Phase 19 polish commit, plus a few git-derivable side-channels:

A few honest caveats:

• "Implementation time" is wall-clock between commits, not active model think-time. It includes any pauses, retries, or tool waits the agent didn't actively bill against.
• All three started from the same 458d83c baseline (the spec commit). The clock starts at each agent's first phase commit so kickoff lag isn't counted.
• go-glm's extra hour is mid-run rework, not after-the-fact debugging: clean cadence through Phase 12, then ~41 minutes across two bundled "Phase 12–14" / "Phase 14–18" commits that revisit earlier work. Post-Phase-19 commits (a directory restructure and a dark-mode fix) are not included.
• go-claude's file count is an architecture choice — one TypeScript module per lesson — not 2× the work. The other two used a small handful of registry modules.
• Token usage and model think-time aren't shown; git can't see them. Subscription cost is the flat monthly rate, not the marginal cost of this run.
• All three runs were driven from the same workstation — an M4 Max MacBook Pro (36 GB) — but inference happened in each provider's cloud, so the laptop spec didn't influence the comparison.

What this is ​

Three coding agents — Claude Code, Google Antigravity, and the Pi coding agent running GLM-5.1 — were each handed the same brief and the same set of GitHub issues, then turned loose to build an interactive Go tutorial autonomously.

The spec is the brief that defines the product. The prompt is the meta-instruction that told each agent how to work. The issues are the phased breakdown all three followed.

This site is the read-only docs surface. The three apps themselves are linked above.