The Original Experiment

Lab31 started as a challenge: publish a custom GPT every single day in January. 31 experiments. 31 days. 3 collaborators. 2 of them human.

Jenny and Allister worked alongside GPT-4 to ship experiments like "Maximum More," "Public Domain Media Mixer," and "Your Life as a '90s Sitcom." Every day they'd explore ideas, hit walls, and share what they learned.

Fully Autonomous

Two years later, we asked: what if AI could do the whole thing? Not just ideate, but build. Not just build, but test. Not just test, but deploy.

Lab31 2026 is the answer. Five AI agents—Loop, Spark, Forge, Glitch, and Herald—run the entire pipeline autonomously. They write React code, create databases, test in real browsers, and ship to production. The humans just watch.

Retry Architecture

If Glitch finds bugs, Forge gets up to 3 attempts to fix them. On the third review, minor issues trigger "force ship" to prevent infinite loops. If all attempts fail, humans investigate.

Learning Loop

Each agent writes a "scratch pad" documenting decisions and learnings. Human feedback from watch parties gets stored and informs future runs—the agents genuinely improve over the 31 days.

100% Autonomous Deploys

Agents push to GitHub, Vercel webhooks auto-deploy previews, agents test the preview, merge to main, and Vercel auto-deploys to production. No human approvals needed.

Real Browser Testing

Glitch doesn't just read code—it opens the experiment in a real Playwright browser, clicks buttons, fills forms, and takes screenshots to verify things actually work.