Give the agent a real browser—not a simulation.
2025-08-20 · genesisA true story, reconstructed from the machine
The browser became a memory.
This is how ThinkRun learned to give AI agents something they never had: eyes, evidence, and a way back to the moment everything went wrong.
Run story
Software used to leave a trail of code. ThinkRun leaves a trail of understanding.
thinkrun.ai / development-session
● LIVE
Execution log
Seeing is useful. Remembering every action is leverage.
session recording + artifactsCapture video, DOM state, screenshots, decisions, and the human correction.
extension · local · cloudHand the evidence back to the agent. Let it understand what to change.
screen recording → proposed fixThe development trace · verified snapshot 2026-07-09
Not a launch story. A learning curve with receipts.
0commits in the verified project historyexact snapshot count
0merged pull requestsunique PR receipts
0days from first commit to snapshotAug 20 → Jul 9
0PRD thread pages woven through memorythe “why” behind git
0memory pages consulted for the timelinefacts, failures, decisions
Weekly commit density
The week the machine caught fire.
Peak · 155 commits · Mar 16 week
Six turns of the wheel
Every version changed what “seeing” meant.
Git recorded what changed. The development memory recorded why. Together they reveal six distinct acts.
Act I · Genesis
01Give the agent a browser.
The first 38 commits established the premise: a browser task system with sessions, screenshots, and a visible trail of execution.
Aug–Nov 202538 commits
02
Act II · Core engine
Make the run observable.
Auth, task execution, video playback, artifacts, before-and-after screenshots, analytics, and real-time task plans turned a command into a session you could inspect.
284 commitsPRD-0024PRD-0026
Act III · Extension + local
03Move into the browser people already use.
The Chrome extension and local-control path collapsed the distance between an agent’s plan and the user’s real tab. This was the peak sprint: 530 commits.
Feb–Mar 2026530 commits155 peak week
04
Act IV · Sharing + MCP
Turn evidence into a language.
Public share pages, embeds, recording feedback, MCP parity, and attach receipts made a browser run portable—something humans and agents could point to together.
280 commitsPRD-0039PRD-0048
Act V · Recording pipeline
05The screen became the prompt.
Video import, large uploads, media derivatives, recovery, and finally in-page screen recording shipped live. Record what happened; let the agent propose what happens next.
309 commitsPRD-0089Live verified
06
Act VI · Distribution
Make the evidence reachable everywhere.
Agent-facing CLI/MCP parity, SEO/AEO, remote-control grants, connector work, and publish-drift guards moved ThinkRun from capability toward distribution.
148 commits18 prerendered routesPRD-0101
“The system kept saying success at the wrong level.”The recurring lesson in ThinkRun’s development memory
The next run starts here
Show it.Don’t explain it.
ThinkRun was built around a simple belief: when an agent can see the actual experience, the distance between feedback and a fix collapses.
›