The divide between platforms over AI agents came into sharper focus, while policy moves around international standards for frontier AI took concrete shape. On the research and practice side, it was a day of reports probing how to train agents and how much LLM reviews can be trusted.

Amazon Blocks Meta's Muse From Purchasing on Its Behalf — Shopify Opens All Stores to It: The Battle Over the "Gateway" to the Agent Economy (Updated From Last Time)

According to ITmedia, Amazon has blocked Muse, Meta's AI agent, from making purchases on its site on users' behalf, citing unauthorized access and security risks. Meta, for its part, is expanding integrations — announcing that every store on Shopify will support purchases made through Muse.

Last time, the story was an "initial pace surpassing ChatGPT" and the discovery of a "critical zero-day." What's new this time is that platform responses have split into two camps: keep control of the gateway to transactions — purchasing — or open it up to AI agents. Amazon is tightening its exclusion of agents from its own site, while Shopify and Meta are clearly taking the opposite tack, embracing them. With AI-agent-driven consumption still in the stage before it takes off in earnest, a battle for dominance over the payment and purchase flow appears to have begun.

OpenAI Proposes International Standards for Frontier AI — "RSI Has Not Happened Yet"

ITmedia reports that OpenAI has proposed that the United States lead the development of international standards for frontier AI. The proposal includes coordination among national AI safety institutes and calls for a framework to prevent uncontrollable autonomous evolution. What stands out is its explicit assessment that "RSI (recursive self-improvement) has not happened yet."

CEO Sam Altman is also scheduled to speak at an open meeting of the UN Security Council. Coming right on the heels of the US and China agreeing to official AI dialogues, discussions over safety institute coordination and standard-setting look set to accelerate. Rushing to build a framework while stating that the risk "hasn't happened yet" can also be read as a position that grounds standardization in preparing for future risks.

Research on Iteratively Training "Category-Specific Experts" for Software Agents and Merging Them Into a Single Model

A paper proposing a new training method for software engineering agents has appeared on Hugging Face Daily Papers ("One to More, More to One: Category-Aware Iterative Expert Training for Software Engineering Agents").

The core of the method is a two-stage process. First, an iterative Refresh–Repair–Expand cycle trains category-specific expert models one after another. Then, label-routed MOPD merges them into a single deployable model. Another distinctive feature is that it does not rely on external solution-trajectory teachers. It achieves 58.04% on Pro-618 and 59.00% on SWE-bench Multilingual, and the model itself has been released on Hugging Face. This approach — bundling finely divided specialties instead of leaning on one giant teacher — is drawing attention as a direction for cutting the training costs of practical agents.

AI Review Quality Should Be Judged by "Evidence," Not "Vibes" — a 7B Fact-Checker Beats a 30B Reviewer

Two reports gauging the reliability of LLM-generated reviews are drawing attention. One is an evaluation by OpenTeams, in which a 7B-parameter fact-checker outperformed a 30B-class LLM reviewer while removing not a single correct claim. The other compares AI code review quality along the axes of "vibes" versus "evidence," and is being discussed on Hacker News.

A small verifier specialized in a narrow task turned out to be more trustworthy than a larger model — results like these look set to serve as counter-evidence against the idea that reviews can simply be left to the biggest model available. As the cost of having humans check generative AI output becomes a real concern, the practical weight of the question of how to design the verification side keeps growing.

First-Ever NetHack Ascension by an LLM Achieved

Hacker News is buzzing over the first-ever ascension (game completion) of the classic roguelike NetHack by an LLM. According to the published dumplog (play record), the feat was achieved under the player name "CodexDelver."

NetHack employs permadeath and is known for demanding an enormous space of item interactions and long-term planning. It is a game where even human players often need a long time before their first clear. That makes this a meaningful win in precisely the areas agent research has struggled with: long-horizon exploration, planning, and risk management. Note that the model and methods used for the achievement cannot be determined from publicly available information at this point.

Robin Williams's Daughter Tells Makers of AI Video to "Have Some Shame" — Backlash Against AI Content Recreating the Deceased

Variety reports that Robin Williams's daughter has objected to fans creating videos that recreate her father with AI, telling them to "Have Some Shame." On Hacker News, the story has gathered more points than any other topic in this roundup.

When it comes to content that uses generative AI to recreate the deceased or real people, how to protect the wishes of bereaved families and control over one's likeness is widely regarded as an area where institutions have yet to catch up with the technology. Each time the voices of those affected surface, platforms and creators are likely to face renewed scrutiny over how they respond.