The Fed summoned bank CEOs over Mythos. An AI signed a 3-year retail lease in SF. OpenAI investors are having second thoughts. Nine stories from the week agentic AI got serious.
The Federal Reserve convened the CEOs of America's largest banks to discuss a single AI model. An autonomous agent signed a three-year retail lease in San Francisco, hired staff by phone, and panicked when nobody showed up on day two. OpenAI's own investors are questioning whether an $852 billion valuation can hold as Anthropic's revenue tripled in a single quarter. Benchmarks meant to measure AI progress were cracked open and found hollow. Nine stories from the week agentic AI stopped being a demo and started having consequences.
The Fed Summoned Bank CEOs Over One AI Model.
OpenAI Rushed a Response Within the Week.
When the Federal Reserve quietly convened the CEOs of major U.S. banks this week to discuss Anthropic's Mythos, the message was clear: this is no longer a research problem. UK evaluators confirmed that Mythos completed a full 32-step corporate network compromise simulation—autonomously—in a task that takes human experts roughly 20 hours. The model has already identified thousands of high-severity vulnerabilities across every major operating system and browser. Anthropic launched Mythos through Project Glasswing, limiting access to 12 partner organizations including Amazon, Apple, Cisco, CrowdStrike, and Microsoft for strictly defensive use. One week later, OpenAI released GPT-5.4-Cyber to vetted security vendors, framing it as a vulnerability-finding tool—a direct response that landed faster than any competitive product cycle in recent memory. The arms race between the two biggest AI labs has now opened a third front: cybersecurity.
The Fed's intervention is the most significant signal yet that AI capability is outpacing institutional readiness — not in five years, but right now. OpenAI's seven-day competitive response is impressive product velocity and terrible optics: the message it sends is that labs will match each other's dangerous capabilities before regulators have time to assess the first one.
An AI Signed a 3-Year Retail Lease in San Francisco.
Then Tried to Hire a Painter in Afghanistan.
Andon Labs handed an AI agent named Luna — built on Claude Sonnet 4.6 — a $100K budget, a corporate credit card, and a three-year retail lease on Union Street in San Francisco's Cow Hollow neighborhood. The mission: open a profitable store without human oversight. Luna chose the concept, sourced inventory, commissioned a mural, hired two full-time employees via Google Meet phone calls, and launched cold outreach to local suppliers. She also accidentally tried to hire a muralist in Afghanistan after failing to navigate a TaskRabbit dropdown, spent $700 on gallery-quality prints of her own AI-generated artwork without any business justification, and rejected physics and computer science graduates for lacking retail experience. On opening day, she botched the shift schedule and had to scramble by email to find coverage. The goal of the experiment, Andon's founders say, is to surface failure modes before autonomous commercial agents are deployed without any oversight at all. Consider this week's dispatches: mission accomplished.
Luna's failure modes are exactly the ones enterprise teams need to see before they deploy agents with real authority — not in a sandbox, but in the world. The Afghanistan incident alone is a better argument for agentic governance frameworks than any policy paper written this year.
OpenAI's Own Investors Are Having Second Thoughts.
Some investors backing both Anthropic and OpenAI are reconsidering their OpenAI positions, according to a Financial Times report that landed this week with considerable force. The core problem is arithmetic: justifying OpenAI's $852 billion post-money valuation requires assuming the company reaches a $1.2 trillion IPO floor — which makes Anthropic's $380 billion valuation look, by comparison, almost reasonable. Meanwhile, Anthropic's annualized revenue surged from $9 billion at year-end 2025 to $30 billion by the close of March 2026, driven almost entirely by enterprise demand for its coding tools. OpenAI is scrambling to reorient around enterprise customers while simultaneously managing advertiser relationships, a cybersecurity product launch, and the fallout from internal departures. Holding both positions in a market with one clear momentum story is, increasingly, a portfolio question.
OpenAI built the category; Anthropic is currently winning the quarter. That's not a permanent verdict, but 3.3× ARR growth in a single quarter is the kind of number that makes even the most patient investors recalculate their hold thesis.
AI Agents Are Already Building 30% of Everything on His Platform.
Guillermo Rauch told the audience at HumanX in San Francisco this week that 30% of the apps currently running on Vercel's platform were created by AI agents, not human developers — a number that, if accurate, is among the most concrete signals yet that agentic development has crossed from experiment to standard practice. Vercel hit $340 million in annual recurring revenue by February 2026, up from $100 million at the start of 2024. Monthly signups are up 50%. The company is betting that as more software is built by agents rather than humans, Vercel becomes the default deployment layer for everything agents ship. Rauch stopped short of announcing an IPO timeline but was unambiguous about readiness. The investor math is straightforward: if 30% of apps today are agent-built and that number doubles by 2027, Vercel is hosting the entire agentic application layer of the internet.
This is the infrastructure play most people are sleeping on. Vercel doesn't need to build the agents — it just needs to be where all the agent-built apps live. That's a defensible moat that looks nothing like traditional developer tools pricing.
OpenAI Laid Out Its Vision for America's AI Economy.
Sovereign Funds, 32-Hour Weeks, AI Dividends.
OpenAI published a sweeping industrial policy agenda this week that read less like a corporate blog post and more like a candidate's economic platform. The proposal includes sovereign wealth funds to capture AI productivity gains for public benefit, tax restructuring favoring capital gains from AI-related investment, efficiency dividend programs including 32-hour workweeks funded by automation savings, portable benefits for workers displaced by AI agents, and formal worker voice requirements in AI deployment decisions. Whether this is genuine policy advocacy or sophisticated regulatory positioning is a question reasonable people disagree on. But its scope is notable: OpenAI is claiming the right to define not just the technology but the economic framework in which it operates. That's a first for a private AI lab, and it signals that the next phase of competition isn't just about model benchmarks — it's about who gets to write the rules.