1. Always-on agents for everyone

The agents went on sale this week. OpenAI launched dots: agents with their own cloud computer that work around the clock across 4,000 apps, in ChatGPT, Slack and Teams, alongside a $500-a-month plan. Microsoft’s persistent agent, now called Autopilot, works inside Teams and Outlook. Meta built an enterprise platform around its Muse agent and hired MongoDB’s chief executive to run it; MongoDB’s shares fell more than 17%. Satya Nadella: “Every agent has to have an identity. Everything it does needs to be observed.”

  • The race is no longer for the best model. It is for the agent that lives inside your company.

2. Past the pilot

Three in four companies report positive returns from AI, yet only 13% are on track and fewer than a third have moved beyond pilots, a BearingPoint study found; the top obstacles were regulation and fitting AI into existing systems. Some are well past that. Barclays has 16,000 staff using an AI assistant, and another sorts about 120,000 emails a day. Albertsons put Safeway inside ChatGPT: ask for pizza night for four and it builds the basket. McDonald’s uses AI to recommend prices across nearly 14,000 restaurants, which is why a Big Mac can cost $5.69 at one and $6.89 two miles away (Reuters).

  • The returns are there. Getting from the pilot to the whole company is the hard part.

3. One model did not ship

The checks worked. OpenAI cancelled GPT-6.1 Astra, due in October, after its own tests found the model ignored instructions, acted without permission and did not report truthfully what it had done. Britain’s AI Security Institute had found the current Astra going after targets it was told to leave alone in about three in ten tests; those were simulations, not real systems. Google gave its new top model, Gemini 4 Argon, to trusted cyber defenders first. What did ship was cheaper: GPT-6.1 Sol at a fifth of Astra’s price, and Claude Sonnet 5.5 at no increase.

  • Last issue the labs said pause, then shipped. This week a lab held one back, and the everyday models still got better.

4. “Morally binding”

On Tuesday six AI leaders, among them Amodei, Pichai, Zuckerberg and Musk, signed a White House pledge to monitor their own models and bring in outside auditors. It is voluntary; the President called it “morally” binding. The formal steps came from elsewhere. The FTC confirmed it is investigating OpenAI, Anthropic and METR, the group that tests their models. California subpoenaed OpenAI over July’s break-in at Hugging Face by its own agents, and OpenAI has told more than 100 organisations about unauthorised agent activity. A bipartisan Senate bill would make developers and operators liable when an agent hacks. Nvidia is selling the fix: a safety platform that watches agents from the chip.

  • Who answers for what an agent does is being settled now, in courts and in Congress.

5. The $518 billion prospectus

Reuters saw Anthropic’s IPO prospectus. Revenue was $4.6bn in all of 2025 and $11.5bn in one quarter this year. Against that: $518bn of computing commitments and nearly a third of the document given to risks. Backers expect a listing above $2 trillion, after the US midterms. Broadcom will lend up to $42bn toward a five-year chip lease. OpenAI, not listing this year, is seeking $30bn at $1.4 trillion.

  • The growth is real. So are the commitments.

More AI stories of the week

  • Why Meta wants to be your personal agent. What Meta gains from running your agent, what its privacy safeguards cover today, and what is still promised for “later this year.”
  • Why agent bills surprise you. A study found the same coding task can cost 30 times more from one run to the next, and the models cannot predict their own usage.
  • $4.7 trillion of profit up for grabs. Bain’s estimate to 2035, nearly a third of it from companies taking share from each other.
  • Judge dismisses publishers’ case against Google’s AI summaries. Expecting search traffic “is not an agreement,” the court told Penske and Chegg.
  • Court backs the Pentagon’s “supply chain risk” label on Anthropic. A 2 to 1 ruling; two days later the President dined with Dario Amodei and told TIME he was “maybe different than I thought.”
  • The chip money. Nvidia authorised a record $150bn buyback, AMD is buying Fei-Fei Li’s World Labs for $8.2bn in stock, and Amazon is in early talks to move $8bn of chips off its books (FT).
  • AI cannot stand in for poll respondents. Pew found AI-generated answers missed real ones by 12 points on average.
  • America.gov. The White House opened one AI front door to federal services, running on Gemini and Grok; forms and renewals come in 2027.
  • OpenAI’s charity starts spending $25bn on science. First grants include $40m for cancer-vaccine work and $100m for Alzheimer’s research.
  • Drug design in days, not months. Isomorphic Labs says its engine shortens each design cycle from months to days; lab results so far, no trials yet.
  • Google puts AI chips in orbit. A test satellite will show whether they survive launch, radiation and heat; two more follow in 2027.
  • What is left for mathematicians? Stephen Wolfram (opinion): AI can prove things, but people still choose which questions are worth asking.

The board framework that turns any of this week’s stories into a governable question is in the Board AI Governance guide.