The old question
“Can an agent click the right thing?”
notenough
CHAPTER 01
GENESIS · First commit
A browser was never just a browser.
To an agent, a browser is a shifting world of pixels, state, intent, and consequence. The first problem was control. The deeper problem was understanding.
enough
The ThinkRun question
“Can it understand what happened?”
CHAPTER 02
THE LOCAL TURN · Extension + bridge
Then it moved into the browser you already use.
Cloud automation learned to meet the messy, signed-in, human web.
The extension, native host, CLI, and local bridge turned the user’s own Chromium into a controlled workspace—scoped to attached tabs, observable, and interruptible.
That changed the product’s center of gravity: from remote browser service to a shared surface where human and agent could hand control back and forth.
CHAPTER 03
THE SYSTEM · From action to evidence
The trace became the product.
A recording stopped being a video file. It became a structured handoff: what was seen, what was done, what broke, and where an agent should begin.
Browser state
The actual page, in the actual session.
Record
Screen, voice, clicks, console, and network.
Synchronize
Events anchored to their exact moment.
Agent reads
Evidence in a form a model can use.
Propose fix
A reviewable next move, inside the workflow.
CHAPTER 04
THE EVIDENCE · Raw history, not mythology
The work left receipts.
The development record is unusually visible. Git history says what landed; PR threads say what survived review; memory says why the direction changed.
1,570
Counted directly from the repository history.
572
Unique PR numbers found in merge subjects.
59
The decision trail behind the code.
323
Days from 2025-08-20 to 2026-07-09.
CHAPTER 05
THE MEMORY · The development loop
Software that remembers why.
The unusual part was not the velocity. It was the feedback loop.
Plans, implementation notes, corrections, deployment gotchas, and product decisions were kept close to the code. Each run could inherit the hard-won context of the runs before it.
✓
Observe the real outcome.
Not merely whether the command returned success.
✓
Preserve the correction.
So a solved mistake does not become a recurring ritual.
✓
Make the next handoff legible.
To the human, the reviewer, and the next agent.
THE NEXT RUN · waiting for input
It is not a finished story. It is a live session.
ThinkRun’s development story ends where every useful recording begins: with someone noticing what should be better, pointing at the evidence, and handing the next move to an agent.
Status: still running