Blog

Notes on recording, AI coding agents, and shipping faster.

• What 331 days of AI-agent-built software actually looks like (with receipts)

  2026-07-17

  ThinkRun has been built almost entirely by AI coding agents since August 2025. Here's the honest version of that story — the silent failures, the 13-round review battles, the same-day rollback — and the evidence gallery where we published 40 agent-written versions of our own history.

• AI debugging when the agent can't see — how a screen recording became a shipped fix

  2026-07-08

  Coding agents can drive a browser now — but only for bugs they can reproduce live. We recorded a real production bug with ThinkRun, handed the structured artifact to Codex, and shipped the fix. Here's the whole loop, including the actual artifact the agent read.

• I let an AI agent set up Search Console — here's what broke

  2026-06-29

  I gave an AI agent my real logged-in browser and told it to verify ThinkRun in Google Search Console and Bing. It mostly worked — but the walls it hit are a tour of how the modern web fights automation, and where handing control back to a human is the right call.

• How to give Claude Code access to your browser

  2026-06-25

  Stop pasting screenshots and explaining what went wrong. Record the browser session once and hand Claude Code a structured artifact it can act on — with your real logged-in session.