Meta, Walmart and Stripe back an open standard for AI agent commerce

Who gets to decide whether the thing on the other end of a transaction is a person or a program? That question sits underneath most of today's news, and I think it matters more than any single model release.

Start with the commerce story. Meta, Walmart, Stripe and Bret Taylor's Sierra Technologies are backing a Personal Agent Protocol, an open standard meant to help businesses tell whether they are dealing with an AI bot or a human customer. Personal agents already book flights, schedule appointments and shop on our behalf, and merchants have mostly been guessing at what they are looking at. A shared protocol turns that guesswork into something negotiable: identity, permissions, liability. I find it telling that the companies writing the standard are the ones who profit most from agents transacting at scale. That doesn't make the standard bad, but whoever defines "trusted agent" will quietly shape who gets access to the marketplace. Watch the governance, not the press release.

Meanwhile, OpenAI is working with Ironclad to train and evaluate agents on complex contracting workflows, with the explicit aim of pushing computer-use capabilities into professional work. Contracts are a smart choice. They are structured enough to measure, consequential enough to matter, and tedious enough that nobody will mourn the handover. By the way, a new survey mapping more than 40 agentic systems across nine task domains found a shared architectural core and, more usefully, persistent failure modes. Put those together and the picture is sobering: we are connecting agents to payments and legal paperwork while the field still hasn't solved how they fail. If you build with agents, design for the failure case first.

Benchmarks have a related credibility problem. GitHub's ReviewBench is supposed to be a shared standard for AI code reviewers, and Copilot wins it. But the team behind the benchmark ran the initial tests for every competing tool themselves, and an independent evaluation paints a different picture. I wouldn't call that scandalous, just predictable. When the referee also owns one of the teams, the scoreboard deserves a second look. Treat vendor-run leaderboards as marketing until someone without a stake reproduces them.

On the model side, Mistral has opened access to Mistral Large 4, its most capable model yet, in public preview, alongside a roadmap. For European teams that want a strong open alternative to the American labs, this is worth testing. Whether it holds up against the frontier models is something your own evaluations will tell you faster than any announcement.

Finally, the quiet one: the General Services Administration finalized a clause on September 28 setting procurement requirements for large language models bought by federal agencies. Procurement rules rarely trend on social media, yet they tend to outlast the models they regulate, and vendors will build to whatever the government's checklist demands. So here is my open question: when the buyer, the standard-setter and the benchmark-runner are increasingly the same few players, who is left to check the work?