92% less time per request, on a slowdown we hid for it From 3,337 to 281 nanoseconds per request, the middle values over 7 paired runs, in a small C program on 30 September 2026. With 95% certainty the cut is between 90.8% and 93.0%.
How it works
Nexus works in five steps: agree, find, fix, prove and report.
First it agrees on the one number that matters and how to measure it. It measures the start several times, to see how much the number moves on its own. Then the agreement is locked, so the goal cannot move later.
Next it measures where the time and the memory really go, before it changes a line. A fix counts only when the old and the new version, run in turns again and again, show the gain every time. Every number in its report must be one it really measured.
Its senses
An AI engineer is only as good as what it can see and do. We built it these tools:
- Eyes, nxvision. So it can see. It opens a real browser, watches every frame move and clicks through like a person.
- Hands, nxterm. It types into real terminals and reads the screen back, colours and all.
- A map, Repomap. It knows where everything is in a codebase before it reads a line.
- A second opinion, nxagent. It asks other AIs to check its work, and they ask it back.
- Maths, nxmath. It computes and checks its maths instead of guessing.