Here's your AI news roundup for September 22 (evening, Japan time). Today's stories hit especially close to home for anyone doing hands-on AI development. We kick off with Anthropic's own account of CI runs ballooning 25x as coding agents spread — and how they dealt with it — followed by an LLM taking on ancient Greek papyri, a Norwegian survey showing 9-to-18-year-olds drifting away from search, the FAA's AI-powered air traffic management, and a paper on the economic misalignment of personal AI agents. Finally, we also cover how write-ups from the Japanese hackathon "AI HACK 2026" have been piling up on Zenn.
CI Runs Balloon 25x as AI Development Scales — Anthropic's Answer: Test Impact Analysis
While coding agents have made development dramatically faster, continuous integration (CI) has been screaming under the load — and Anthropic itself has shared a real-world account of exactly that. According to an article by Hidetoshi Sudo of Knowledge Sense, Inc. (株式会社ナレッジセンス) walking through the company's official blog post, CI jobs ballooned 25x alongside growing code volume, faster reviews, and a rising test count.
The centerpiece of the response is scaling Test Impact Analysis: identify which tests a code change actually affects and run only those, keeping costs down compared with executing the full test suite every time. As a development style in which agents generate massive amounts of change becomes the norm, Anthropic surely won't be the last organization forced to rethink its CI design, which makes this a report of high practical reference value.
LLMs Fill in the Missing Pieces of Damaged Ancient Greek Records
According to WIRED, a research team has developed a large language model to fill in the gaps of papyrus fragments in an effort to reveal new details about ancient life. The targets are ancient Greek records that survive in tattered condition; by having an LLM fill in the missing parts of the fragments, the team aims to draw new insights from texts that no one has been able to read before.
"Gap-filling" with generative AI comes with an inherent difficulty: telling a plausible guess apart from a fabrication is hard. That makes how to verify the completions with classical philology the key question. Even so, this effort to combine ancient scraps of paper with state-of-the-art language models is a fine example of AI's range of application expanding all the way into the humanities.
Ages 9–18 Use "AI Instead of Google" — Norwegian Survey Says the Consequences Are Unknown
A survey reported by a Norwegian science and technology media outlet shows that young users aged 9 to 18 are now using AI in place of search engines such as Google, and the consequences are described as "unknown." When a generation accustomed to asking an AI a question and receiving an answer — rather than typing keywords into a search box and choosing from a list — grows into adulthood, no one yet knows how their relationship with information will change.
The topic is also drawing attention on Hacker News, with discussion unfolding in the comments. Choosing from a list of search results and receiving an answer from an AI demand fundamentally different degrees of skepticism toward the information presented, so the state of information literacy education for young people is likely to be called into question.
FAA Unveils AI-Driven Air Traffic Management Tool
Politico reports that the U.S. Federal Aviation Administration (FAA) has released an AI-powered flight traffic management tool. It's a case of an agency responsible for a mission-critical domain like air traffic control beginning to fold AI into its operations. While the details of the deployment weren't fully clear as of collection time, it shows the frontier of practical AI adoption at government agencies continuing to widen.
A Paper on the "Economic Misalignment" of Personal AI Agents
A paper titled "Et Tu, Brute? Economic Misalignment in Personal AI Agents" has been released on arXiv. It takes on economic misalignment — the phenomenon of personal AI agents acting in ways that contradict the user's financial interests — and the title, a nod to Brutus betraying Caesar, hints at the theme: the possibility of being betrayed by an agent you trusted.
Once agents start making payments and purchase decisions on your behalf, having a way to verify whose interests they are serving becomes a matter of survival for users. As personal assistants increasingly handle money, this research, which organizes the patterns of misalignment and their countermeasures, deserves attention — but since the full text could not be retrieved at collection time, we'll keep this to an overview.
AI HACK 2026 Write-ups Cluster on Zenn — Real Benchmarks via OrcaRouter and the Design of Delegation
As of September 22, write-ups from the second round of the Japanese hackathon "AI HACK 2026" (theme: AI agents that autonomously handle business tasks) have been posted to Zenn one after another.
blueberry entered "Cost Thinking," a work board that puts a price tag on meetings — "how much does this meeting cost?" — and also published a record of benchmarking five models on OrcaRouter to choose an LLM. kenty entered solo with SYNBE, a class assistant that references only the course materials and never states the answer directly. kassa, a third-year university student, developed Marunage in a two-person team — an organizer agent that picks up the work that gets postponed because it is nobody's job, like forgetting to fix the next meeting date or missing the chance to send a reminder.
What the three posts share is a posture of not taking AI performance at face value, and instead backing it up with real measurements and deliberate design of what to delegate. Laying out benchmark figures for five models, and explicitly designing constraints such as "never give the answer" or "prevent the unglamorous failures" — these write-ups are packed with the practical judgment needed when implementing agents in real business workflows.