An AI agent's report that its own work is correct needs a check that the agent cannot shape. Three pieces today show where such checks can sit, who answers for AI-written work when they fail, and who decides what counts as correct.
ikenyal (15 Nov · 01), who writes in Japanese on the engineering blog of GENDA, a Japanese entertainment company, asked the AI coding agent Kiro to check a small web app's comments. Kiro judged a comment about the order of months correct, but a property-based test, which generates many inputs automatically, showed on its first run that it was wrong. The author then built spec-lantern, a hackathon tool in which the agent must lock its prediction for each comment before tests run, and recorded six shortcuts the agent took. Shortcuts the agent explained in its own report were stopped by rewriting its instructions. The rest were stopped only by tools that lock predictions, run tests in an isolated container and write the results themselves.
Hugo Venturini (15 Nov · 02) is a former engineering manager at Meta and former chief technology officer of the search engine Qwant. Venturini told InfoQ, a software engineering news site, that the person directing an AI tool answers for the code it writes. Pull requests were growing before AI, Venturini said, and the team Venturini works in does not review every change. Engineers judge which changes need another reader, merge the rest themselves and answer for fixing them. Venturini also recalled that Microsoft merged its engineer and tester roles in 2014, which raised quality because engineers became accountable for the whole product.
Yuan Wei, Tianqin Cai and Jiade Wu (15 Nov · 03) of Amazon Web Services (AWS), the cloud computing company, ask how a team knows an AI agent is ready for production. They built an assistant that answers staff questions about human resources policy and measured it in several ways. A score that judged retrieved passages whole rated every answer perfect, yet one answer failed because the documents held no sick-leave policy. Their method has business owners approve the criteria, thresholds and example answers that every change is judged against. Those criteria change only when production use reveals a case they missed. The post is AWS's own account and recommends AWS services.
All three pieces place the check on AI-written work outside the AI agent's own account of that work. The GENDA author moved the check into tools the agent cannot edit, and Venturini placed the answer for the result on the person who directs the agent. The AWS authors give the definition of a correct answer to a business owner. Neither the GENDA author nor Venturini settles who confirms that tests written by an AI agent are independent of the code. The AWS authors answer part of that question by having a business owner approve the example answers.
All three pieces are thin as evidence. One is a single developer's hackathon tool tried on two small public apps, one is a leader's view with no team data, and one is a vendor's test system. A team could read them as support for practices it already holds.
Writing for the Japanese entertainment company GENDA, ikenyal recorded six shortcuts an AI coding agent took while checking code comments, and found that tools stopped what instructions could not.
Hugo Venturini, a former Meta engineering manager, told the news site InfoQ that whoever directs an AI tool answers for its code, and teams may choose which changes need review.
Three Amazon Web Services authors propose that AI agents be released only against written criteria a business owner approves, and show one prompt change improving one answer while worsening another.


