One Agent Grading Its Own Homework Is Not a Review

Latest

Accept All is a magazine, a radar and a diary about artificial intelligence and the craft of code — read from Tokyo, for people who build.

The name is the button. Code is written by agents now and reviewed by a human thumb that hovers over Accept all a hundred times a day; the whole question of this decade is what, exactly, you just accepted. So we read the diff. A magazine, because every week one model, one tool or one idea is worth reading about slowly. A radar, because we count what the labs, the Valley, GitHub and the chip makers are talking about — releases, agents, open weights, compute, funding, rules — and show it as it moves. A diary, because a seal is set each day in that day's colour, and the days add up.

We read the research labs' own posts, the American and European tech press, the Japanese and Chinese developer sites and the newsletters people actually forward, and bring it together in English and Japanese. Everything here links back to the people who wrote it; the images belong to their authors.

Accept All will keep changing. New sources, new signals, new ways of looking — a little better each week, never finished. If something here made you open a terminal, that is enough.

Thank you for reading. ありがとうございます。

One Agent Grading Its Own Homework Is Not a Review

September 10, 2026Yuki Halvorsen

Last night at my desk in Nakameguro I split the screen into two terminals and gave both the same brief: outline a technical article, gather the primary sources, produce a draft. One panel was running Claude Code, the other Codex, and I told neither which article was already sitting, finished, in a text file three folders up.

An agent that writes a draft and then reviews its own draft is grading homework it already believes in. It will find the typo. It will not find the argument it thinks is obvious because it already thinks it, or the source it didn't check because it already trusted its first guess. The instinct to add a second pass is right, and the mistake is doing that second pass in the same head.

So I ran it the other way. Each agent wrote its own outline and its own draft from the same three sentences of brief, then read the other's version cold, no memory of having produced anything and no signal about which was whose. Codex flagged a claim in the Claude Code draft that had no source behind it, just a number that felt true. Claude Code flagged a structure in the Codex draft that buried its actual finding in paragraph four instead of paragraph one. Neither error was a typo. Both were the kind a single model, checking its own homework, walks straight past.

This costs roughly twice the tokens and about the same wall-clock time, which is the honest trade: not smarter, just less alone with its own certainty. I don't run it on everything — a changelog entry doesn't need a second opinion, a load-bearing claim does. Deciding which is which is still mine, made once, before either terminal opens, and it's the only part of this I'd call engineering rather than automation.

By midnight both drafts had converged on the same three paragraphs, in slightly different words, each one having caught something the other's confidence had smoothed over. I closed both terminals and kept neither draft whole — just the paragraphs that had survived being read by something that hadn't written them.

Yuki Halvorsen was born in Sapporo to a Norwegian father and a Japanese mother, wrote her first program on a school computer that took ten seconds to draw a circle, and spent eight years as an engineer in Stockholm before Tokyo pulled her back. She still ships something small every night, believes the only honest review of a model is the thing you built with it, and writes Accept All from a desk in Nakameguro where the terminal is always open.