T3 Chat builder Julius says his agent merged almost 400 pull requests in 30 hours to speed up the web app
For the past 30 hours my agent has merged almost 400 PRs focused on web perf: • 86% fewer renders when another thread streams • up to 35% less JS CPU • ~10% smaller heap Mobile next. Would be a shame if I had to freeze merges…
Context
A single developer pointing an agent at performance work for more than a day shows how agents are changing the volume of code a small team ships.
This is a first hand post from Julius, who says he builds T3 Chat and T3 Code. The figures are his own and unverified, and he did not say which agent he used or how the changes were reviewed.
Deep dive
Early on October 18, Julius, whose profile says he is building T3 Chat and T3 Code and contributes to the open source projects tRPC and TanStack, posted that for the past 30 hours his agent has merged almost 400 pull requests focused on web performance.
He listed three results: 86% fewer renders when another chat thread is streaming, up to 35% less JavaScript CPU use, and a heap about 10% smaller. He wrote that mobile is next and joked that it would be a shame if he had to freeze merges.
The joke refers to his own earlier post, which he quoted, saying he thought he would start a merge freeze from Friday until he has a Duo in his hands. He did not explain what the Duo is.
At that pace, the agent merged a pull request roughly every four or five minutes. Julius did not say whether the pull requests were reviewed by people, how the improvements were measured, or whether the agent ran alone or as several agents in parallel. Many small automated changes can each be easy to check, but together they can also make it harder to trace the cause of a later bug.
The post did not include benchmarks or links to the pull requests.
What to expect next
Watch whether Julius shares the measurements or the agent setup behind the 400 pull requests, and whether users notice the speed changes in T3 Chat.
Also watch the mobile work he says is next. Developers are increasingly letting agents run for long stretches on narrow goals such as performance, and how teams review that much automated code is an open question.
