AI Coding Weekly

Theo turns T3 Code's worst latency cases into a hill climb

Bad mobile internet plus very long threads was the edge case, so he made it something an agent could score itself against.

Theo turned his worst performance cases into a hill climb running against a set of fake devices and iOS simulators, and posted that it "worked great". The target was a specific edge, bad mobile internet combined with really really long threads, which he describes as a less than ideal experience even though most of T3 Code felt fast. He credited Fable 5.2 and said he was up until 6am on it.

Theo - t3.gg
@theo
X
I turned all my extreme cases into a hill climb with a bunch of fake devices and iOS simulators.
Nov 12, 2026 · View on X

The post follows his own earlier complaint that "T3 Code is still too slow", which is about as blunt a problem statement as a product owner gets to write about their own app.

Theo - t3.gg
@theo
X
T3 Code is still too slow. Working on it!
Nov 12, 2026 · View on X

Why the harness is the interesting part

The copyable move here is not the fix, it is the setup. "Make this faster" is the kind of instruction an agent will happily act on and then tell you it succeeded, because nothing in the loop contradicts it. A hill climb, meaning a scored loop where each change is measured and kept only if the number improves, gives the agent something it cannot talk its way past. Simulated devices and iOS simulators are what make the score mean something for the case Theo actually cared about, rather than for the median request on a developer laptop on good wifi.

That is also why the edge case framing matters. The failures were not spread evenly across the app. They clustered where network conditions were bad and thread length was extreme, the exact combination that a casual local test never reproduces. Building the bad conditions into the harness is how those cases stop being anecdotes from users and start being a number.

For working engineers, this is one developer's result on one app, not a benchmark, and Theo has not published latency figures for before and after. What he has published is the shape of the approach. If you have performance complaints that only show up in conditions you cannot easily sit in, the question worth asking is whether you can fake those conditions well enough to score against them, and then let the agent optimize the score instead of your description of it.

Have the good opinion, not the loud one

One email a day. Under three minutes. What working engineers found out about coding with AI.

Get the next one by email

Coding with AI, read daily so you do not have to. The experiments, the receipts and the arguments from engineers shipping real software. Not a changelog.