The config said yes.
The library said no.
A ClearAuth integration silently dropped email/password configuration, causing a multi-day signup outage. Lesson: test the boundary, not the label.
The development story of ThinkRun
So we taught agents to see it, replay it, remember it—and eventually, act inside it.
In August 2025, the first commit began with a simple mismatch: an AI could change a product, but the truth of that product lived somewhere else—in a browser tab, behind pixels, clicks, requests and state.
The screen was where the evidence was. The agent was blind to it.
NO CONSOLE · NO NETWORK · NO CONTEXT
The answer wasn’t another screenshot. It was a flight recorder for the web: the visible moment and the invisible machinery, time-synced into evidence an agent could actually use.
The first browser automation service proved the loop: give an agent a task, let it navigate, preserve what happened.
Verified snapshot: 2026-07-09 · source: git history + memory wiki
Illustrative cadence · not a week-by-week data chart
March 16–22, 2026 was the densest week in the verified history: 155 commits. By then ThinkRun had grown a Chrome extension, a local bridge, recording, and shareable sessions.
A commit graph shows velocity. Memory shows consequence: auth that looked configured but wasn’t, uploads broken by an allowlist, an install command that simply did not install. The incidents became architecture.
A ClearAuth integration silently dropped email/password configuration, causing a multi-day signup outage. Lesson: test the boundary, not the label.
Recording uploads failed when the R2 CORS allowlist drifted from the live origin. Lesson: infrastructure outside git is still product code.
A missing shared package made the homepage’s install path a total outage. Lesson: smoke-test the promise, not just the package contents.
Daily logs, PRD threads, incidents, corrections and architectural decisions became a development memory wiki. Not a victory scrapbook—a mechanism for carrying hard-won context into the next decision.
“Verify the artifact, not the claim.”
ThinkRun began because agents could build what they could not inspect. It became a way to capture the whole moment—screen, voice, clicks, console and network—and hand it back as working context.
The browser was the missing context. ThinkRun made it legible to agents.Built from 323 days of verified project history · no external assets