41 Rounds, Zero Manual Edits: The Agent's Scaffold Starts Rewriting Itself

First, two things. One comes from virgilxbt's original tweet, saying a Chinese lab published a paper titled 'The End of Harness Engineering.' The other comes from slash1s's original tweet: a 3D office where every desk has an AI coworker, and finance Jessica will speak up and roast people.

A little cold water first. For the paper, no public link is visible yet. In the replies, some users say they can't find a paper with that title, and others say it sounds like Self-Harness and Google's OO-Harness. The original tweet only provided screenshots and a video. Below I'll break it down according to the original tweet's claims, and set the conclusion aside for now.

The original tweet boils the paper's core into one sentence: the file that organizes the model should be written by the loop that runs it, not by the engineer who builds it. A manual harness is a guess with a commit message; a self-modifying harness is a file that scores itself every round.

The loop is almost boring. An execution report goes in, eight candidate edits come out, they're scored on held-out questions (a scoring set held out from optimization), the winner is kept, and the process repeats. The solver stays frozen throughout. After 41 rounds, the held-out score goes from 0.31 to 0.64, with no human touching the file.

Humans only write the scoring set. The proposer can't see it, candidates don't optimize against it, and it only scores each candidate. The work shifts from arranging calls and tools to deciding what counts as correct, then reading the scores. The harness becomes disposable; the scoring set becomes the work.

The original tweet's author says he ran exactly this loop himself: one harness.py, 300 Kimi K3 workers, 37 rewrites across 41 rounds, zero manual edits.

There's a point here that's easy for the headline to obscure. A self-modifying harness is not maintenance-free. If the scoring set is written wrong, the loop will steadily optimize in the wrong direction. The engineer's job narrows, and also gets harder. Writing a scoring set is closer to product judgment than writing a scaffold, and harder to outsource.

Some replies also pull the question back to control: labs want models to do more, but humans don't work unsupervised, so why can bots? That objection is closer to the engineering reality than the headline.

Now the office. slash1s walks into a 3D office: reception, finance, copier, whiteboard. WASD to walk, E to talk. He finds finance Jessica and says he trades crypto. Jessica replies: 'Your relationship with volatility must be very interesting—don't expense your losses to the company.'

Tavus's PAL connects to voice in real time and has a face. The original tweet says you can make one in a few minutes. Jesse Rowe had Claude build this office for Tavus PALs; the original video is here.

slash1s's original tweet and Jesse Rowe's original tweet are both here.

This one is lighter, but it points to the same thing as the paper: the Agent is no longer just text in a chat box; it starts to have an environment, characters, scoring, and feedback. One uses a scoring set to force the harness to evolve; the other uses 3D space and voice to force the character to be believable.

The difference is verification. With the office demo, you can watch the video and judge whether it's interesting; for the paper, we have to wait for public materials. For now, remember two verifiable points: 41 rounds, 0.31 to 0.64, zero manual edits; and Jessica roasts the trader. The former determines how engineering changes; the latter determines whether you'll press E at your desk tomorrow.