Builds

Interactive proofs, dashboards, and case studies that make a new AI capability obvious: what changed, what held up in practice, and the patterns worth borrowing.

Interactive⚡ 1 session

Browser Task Replay

What this proves

A failing agent run is a logbook, not a verdict. The replay turns a synthetic trace into a row you can fix on Monday.

Step through a bundled synthetic agent trace: click, wait, assertion, and the row where stated intent did not match what the agent actually did.

google-io-2026chromeagent-traces

Get new builds, breakdowns, and useful AI updates.