Give the agent a window.
The first version was direct: create a browser, give an agent commands, return what happened. Navigation. Clicks. Screenshots. A small mechanical vocabulary for reaching through the glass.
This is the story of how a remote-control experiment became a recorder, a debugger, a handoff, and a shared language between people and agents — told from the repository that watched it happen.
Run the historyAt first, the browser was just somewhere an agent could go.
The first version was direct: create a browser, give an agent commands, return what happened. Navigation. Clicks. Screenshots. A small mechanical vocabulary for reaching through the glass.
A demo clicks once. A product survives the second click, the cold start, the half-created session, the lost worker, the retry. Auth, resilient tasks, structured errors, and video playback turned browser control into infrastructure.
The browser left the cloud and met the one already open on your desk. A Chrome extension, native bridge, tab ownership, local sessions, and obstacle recovery made “use my browser” a real sentence.
A browser session stopped being private state and became an artifact: replayable, embeddable, linkable, readable as Markdown or JSON, and callable from an agent through MCP. The run could finally travel.
Video alone shows the symptom. ThinkRun learned to keep the evidence around it: voice, clicks, console, network, DOM, timeline, and analysis. The web recorder and Activity Hub turned “look at this” into agent-ready context.
SEO, skills, publish-drift guards, remote control grants, and hosted connectors brought the system to more agents in more places. The product was no longer one browser service. It was a route into the browser from wherever work begins.
Daily logs became memory pages. PRDs became threads. Commits, reviews, failures, and corrections became a wiki that can explain not just what changed, but why. This page is one result of that memory looking back at itself.
Each bar is an exact unique merged-PR count for its phase. Together they reconcile to the repository total — no estimates, no decorative data.
An earlier timeline draft called stale estimates verified and invented a plausible explanation for arithmetic that did not reconcile. The memory system flagged it. The history was recomputed against raw Git, strict boundaries, and exact PR receipts. The correction became part of the record.
Verify the artifact, not the claim. A system that remembers only its successes is a brochure. A system that remembers its corrections can improve.
People show. Agents inspect. The browser keeps the evidence. And the next person — human or machine — starts from what really happened, not from a retelling.