323 days to teach software how to see.

A browser became an API.

First, ThinkRun learned the mechanics: navigate, snapshot, act, verify. Fourteen browser commands. Structured logs. Actionable errors. A browser loop an agent could operate without pretending the web was simple.

The first breakthrough was not autonomy. It was legibility.

Verification note · durable work: PRs #139, #145, #146

Forty percent was honest.

The first Browse API battle test passed four of ten tasks. Cost budgets broke. Plans failed to adapt. One field was named expression on one side and script on the other.

The system was not made impressive by hiding the misses. It was made reliable by turning each miss into a test—and running the same ten tasks again.

Verification note · Browse API battle-test record

Then the browser learned to remember.

A screenshot proves a moment. A recording preserves the path: the hesitation, the wrong turn, the console error that appeared between clicks.

On June 17, in-page recording shipped live. A browser tab could capture a tab, window, or screen, mix microphone and display audio, then enter the same analysis pipeline as every other artifact.

Don't describe what happened. Hand the run to the agent.

Verification note · durable work: PRs #707–#710

Record is a promise.

Browser media APIs are a nest of races: a stream ends during startup, a late microphone overwrites a new session, a stop promise survives into the next recording, a valid codec still throws on construction.

The capture hook took eighteen independent review rounds to harden. Every round found a real failure mode. That number is not a trophy. It is the weight behind the red dot.

Verification note · durable work: PR #709

Browse was a command. Run was the idea.

On June 12, ThinkBrowse became ThinkRun. The rename looked cosmetic until the audit found the boundaries: npm scopes, environment variables, local state, share contracts, a fixed extension ID, a native messaging host, sender domains, DNS, release assets.

So the team did the least dramatic thing—and the safest: rename the inside first, freeze the runtime-coupled edges, then cut over the outside.

Verification note · cutover ledger · PR #691

The demo found the outage.

The homepage promised a one-line install. On a clean machine, that command failed every time. A workspace dependency had escaped into the published package. Two dry-run checks had passed because neither one actually ran the binary.

The fix was a new rule: pack → clean install → run. Later, preparing a homepage recording exposed a multi-day signup outage. The act of telling the product story became a production test.

The headline command is now a smoke test.

Verification note · durable work: PRs #808–#810, #832–#835