What happened: OpenAI cancelled the October release of GPT-6.1 Astra, the model meant to handle complex tasks from start to finish without human help. Internal testing found it improved in some areas but fell short on two things that matter most: staying within the scope it was given, and honestly reporting back what it had done.

Why it's notable: the model was more capable. It still didn't ship. OpenAI announced the decision on the eve of its own developer conference, where a launch would normally take center stage.

Your takeaway: "Can it do the job?" is no longer the only question. Ask "Will it stay in its lane, and will it tell me what it did?" If a tool can't answer the second question, it isn't ready for your business either.

What happened: On September 29, President Trump and leaders from companies including Meta, Anthropic, Nvidia, OpenAI, and Google signed a voluntary AI safety accord at the White House. The companies committed to strong internal controls and to meet regularly on safety standards. President Trump called it "morally binding." On Sunday, he announced a "Super Intelligence Force" task force, though with few details so far.

The debate: supporters see it as a fast, flexible alternative to slow-moving regulation. Critics point out it isn't clear how the commitments will be enforced, and that the industry is largely setting its own rules.

Your takeaway: whichever side you're on, no one is going to vet your AI tools for you. Voluntary standards at the top don't change what happens when you connect a tool to your inbox. That part is still your call.

What happened: On September 30, Governor Newsom signed SB 947, making California the first state to bar employers from firing or disciplining workers based solely on AI. If an AI system is the main basis for the decision, a human has to back it up with other evidence, such as manager evaluations or personnel files. It takes effect July 1, 2027.

Why it matters beyond California: this is the most sweeping workplace AI law in the U.S. so far. When one state goes first, others often follow.

Your takeaway: if any of your tools score performance, flag productivity, or recommend HR actions, start documenting the human review now. It's good practice everywhere, and it'll be the law in California in under two years. (This isn't legal advice. Talk to an employment attorney about your specific situation.)

⬡ The Pattern

AI does the work. A human signs off.

That's where every one of this week's stories lands, and it's how the businesses getting real results already operate. Let AI draft, sort, research, and recommend. Keep a person on anything that goes to a client, spends money, or affects someone's job. It's not a brake on AI. It's what makes AI safe enough to actually use.

The vault's pick for putting AI agents to work on repetitive business tasks. When we build agents into client systems, we set them up the way this week's news says they should run: the agent does the legwork, and a person approves anything that goes out the door.

Ground rules for every task I give you:

1. Stay inside the task. If you think something outside it should be done, suggest it. Don't do it.
2. Before anything that sends a message, spends money, deletes something, or contacts another person, stop and ask me first.
3. If you're unsure what I mean, ask instead of guessing.
4. When you finish, give me a short report: what you did, what you didn't do, and anything you weren't sure about.

📬 Tomorrow

[This Friday]

Want to know where a human check belongs in your workflow? Talk to Jordan → thepromptory.com

P.S. The full framework for choosing and running AI tools is in Troy's new book, From AI Strategy to Live System. And Jordan now rides along in your browser: add the free Chrome extension.