How Helix is built — the measurements, the bugs worth writing down, and the design decisions behind an agent that runs on the computer you are sitting at.
The phone could reach ten things on a machine that served 125. The fix was to stop adding verbs — and then to find out the app had been edited in a generated file all along.
Read →A phone watching a task would freeze mid-run and never recover. The agent was fine — the events describing it were being thrown away by a single early return.
Read →Helix remembers in four places. Two of them were asked one question, once, before the task began — and never again. Plus a dedupe bug that deleted facts.
Read →"It feels slow" and "where the time goes" turned out to be different stories. The reasoning was never the bottleneck — blocking timeouts were.
Read →querySelectorAll is a lie on a page built from components. We were reading 58% of a document and reporting it as all of it.
There was a planning layer, and almost none of it was connected to the run. Seven things it decided that nothing downstream ever read.
Read →An agent on your real machine cannot take a guess back. So the interesting design question is when it should stop and ask.
Read →