Every few months, a week comes along where you can feel the pieces snapping into place. This was one of those weeks.

The agentic commerce conversation has been simmering for months. We have written about the trust gap, the standards race, and what happens when agents don't click ads. But this week, the infrastructure layer started hardening. Card networks published competing frameworks. Google formalised an open protocol. Stripe shipped the billing plumbing. And OpenAI released a model that can operate your computer end to end.

The biggest story of the week was Visa and Mastercard jockeying to set agentic commerce standards, each locking arms with big tech partners and digital payment players to shape the ecosystem before it solidifies.

Meanwhile, Splitit announced its backing of Google's Universal Commerce Protocol (UCP), an open standard providing building blocks for AI agents to complete purchases on behalf of consumers. Google is positioning UCP as the HTTP of agentic commerce: a shared language that any agent, any merchant, and any payment rail can speak.

On the startup side, Circuit & Chisel, funded by Stripe and Samsung, is building the navigation tools that AI agents need to traverse the web autonomously. Think of it as the browser engine for agents that never open a browser.

We explored this exact dynamic in our analysis of The Agentic Commerce Standards Race. The thesis holds: whoever sets the rules sets the market. This week, the rule-writing accelerated.

It is one thing to publish a standard. It is another to watch an agent order your groceries.

DoorDash and Uber are now testing true agentic ordering inside Google Gemini. Not chatbot-assisted browsing. Not "here are three options, pick one." Full, multi-step ordering where the AI handles everything from restaurant selection to payment, without the user touching a keyboard.

Retail TouchPoints also published a practical guide on 7 ways retailers should prepare their sites for AI-powered buying, noting that OpenAI's assistant can already complete entire purchases autonomously and Amazon's "Buy for Me" agent is expanding its reach.

We wrote about what happens When the Agent Becomes the Customer. That future arrived this week in food delivery.

Stripe introduced new AI-focused metering and billing capabilities inside Stripe Billing, giving software companies infrastructure to charge for AI consumption the way cloud providers charge for compute: by the unit, in real time.

The update lets developers send granular usage data including tokens processed, model API calls, and agent tasks. It is the billing plumbing that agentic commerce needs to function at scale.

In the funding corner, Dutch cloud-native payment processor Silverflow raised $40 million in Series B funding led by Picus Capital. And UK startup Vivox AI raised £1.3 million to build regulator-ready AI agents for AML, KYB/KYC, and financial crime compliance.

OpenAI launched GPT-5.4, and it is not just another benchmark bump. This is OpenAI's first model with native computer use capabilities, meaning it can operate a computer on your behalf and complete tasks across different applications. The Verge called it "a big step toward autonomous agents."

The same week, OpenAI shipped ChatGPT for Excel with financial data integrations and launched Codex Security in research preview, an application security agent that analyses codebases, validates vulnerabilities, and proposes fixes.

Not to be outdone, Anthropic's Claude Code Security debut shook cybersecurity stocks and intensified competition in application security testing.

We broke down this rivalry in GPT-5.4 vs Claude: The Agent Layer War Just Went Live. The battleground is no longer who has the best chatbot. It is who owns the agent layer.

A few pieces from the research world worth your attention this week.

Researchers at Sapienza University of Rome discovered that when language models hallucinate, they leave measurable traces in their own computations, a pattern they call "spilled energy." The method is training-free and generalises better than previous approaches. If it scales, hallucination detection gets significantly cheaper.

Meta FAIR and NYU researchers found that LLM text data is drying up, but unlabelled video may be the next massive training frontier, challenging common assumptions about how multimodal models should be built.

Liquid AI released LocalCowork, a privacy-first agent workflow system powered by their LFM2-24B model, running entirely on-device via Model Context Protocol (MCP). No API calls, no data egress. Enterprise privacy without the cloud dependency.

And Andrej Karpathy open-sourced Autoresearch, a system where AI agents conduct ML research automatically on a single GPU. The implications for research velocity are significant.

Anthropic's enterprise business is reaching escape velocity, according to Stratechery, with revenue skyrocketing and agents dramatically increasing demand for Nvidia chips.

The SaaStr/20VC podcast dropped a headline that landed hard: OpenAI raised $110 billion privately (four times the largest IPO in history), and Block cut 40 percent of its workforce. We covered Block's story in depth in Block Cut 4,000 Jobs and Blamed AI. Here's What Actually Happened.

And in the spirit of "you don't need a massive team anymore," a solo developer in Taiwan built four AI agents that run his entire company on Google Gemini's free tier. Content, sales, security, and ops. Monthly LLM cost: $0.

This week we also published The Live-Service Graveyard, an opinion piece on how the gaming industry keeps burning billions and burying careers on live-service bets that were dead on arrival. A departure from our usual payments and AI coverage, but a story about what happens when an industry mistakes recurring revenue for guaranteed revenue.

What story from this week landed hardest for you? Hit reply and let us know.