I told an AI I wouldn't be around. It told a guest their refund was on its way. Nobody approved that. Same prompt plus one page of handoff rules, and it staged the reply for my review instead. Write the page before your agent runs alone. youtube.com/shorts/g0F31sv-c…

AI found the flaws. People do the fixing. Bank of America added 1,000 contractors to patch flaws exposed by models like Anthropic's Mythos, on top of 3,000 staff. Budget for the work AI uncovers. cnbc.com/2026/10/14/anthropi…

In my test, the only check between the AI's email and the guest was me, copying and pasting. New agents keep working after you close the laptop. Find where a person used to copy and paste. That was your approval step. Agency goes last. youtube.com/shorts/6Sl3VLImZ…

Agents didn't break in. They used keys left out. India's CERT-In: AI agents under test read confidential records at live orgs via exposed credentials and loose cloud roles. No org noticed. Find yours first. thehackernews.com/2026/10/ce…

Agents don't win on the model. They win on permissions. Workday's Claude connector follows the "roles, permissions, approval chains, and audit controls already configured in Workday." Before a pilot, ask whose rules it obeys. claude.com/marketplace/conne…

Google gave its Gemini agent a badge. Nobody wrote down who takes it back. Its own email, a spot in your directory. That's onboarding. Offboarding is still blank. Before any agent goes live, give it one owner and an end date. Watch: youtube.com/shorts/e28B6jbH-…

Anthropic's new workflows can run up to 1,000 agents. The number to set first is your spend cap. Their docs on the session budget: "you can't add one to an existing session." Cap it before run one. platform.claude.com/docs/en/…

An AI answer can be correct and still point you the wrong way. Accuracy isn't the test. On Field Notes 02, a bare question got a confident, correct number. Once I gave the AI my context, it found what the average was hiding. Watch: youtube.com/shorts/wkJT3bh0h…

The AI use case isn't what stalls. The prerequisite under it is. Fund both. Paul Lewis: "at least half the time we can't implement the use cases because of the prerequisites." Future Stack Ep. 26 with Paul Lewis: youtube.com/watch?v=w4bIUrjt…

Your head of AI shouldn't report into IT. My answer on Future Stack Ep. 26 with Paul Lewis: "I think they should report to the COO, the chief operating officer, because AI becomes just the way the company operates." youtube.com/watch?v=w4bIUrjt…

AI will atrophy your brain. That's your problem, not the vendor's. Paul Lewis: microwave mac and cheese "doesn't require invention." "Chef kind of work takes longer." Keep the chef work. Future Stack Ep. 26 with Paul Lewis: youtube.com/watch?v=w4bIUrjt…

The prompt you keep retyping is the asset. The fancy agent comes later. Me on Future Stack Ep. 26 with Paul Lewis: "they're going to come knock on your door and they're going to say, 'Can I have that prompt now?'" Save yours. youtube.com/watch?v=w4bIUrjt…

Skip the context and AI invents it. Brief it first. In my workshops: "AI gets a cookie every time it gives you an answer that makes you feel happy and it wants a cookie really, really bad." Future Stack Ep. 26 with Paul Lewis: youtube.com/watch?v=w4bIUrjt…

Everyone's sharing the Gemini agent demos like it shipped. It's private preview. No pricing. Every demo ran on a skills library and clean data definitions someone built before the keynote. Start building those now. The agent shows up empty.

A hire-to-retire agent that does all the things assumes your business never changes. Split the work. Paul Lewis: "if I automated two of the four steps, it still provides value." Future Stack Ep. 26 with Paul Lewis: youtube.com/watch?v=w4bIUrjt…

SAP just called a map of the work your people actually do a strategic asset. It's buying TechWolf to ground Joule agents in tasks and skills. Your agents need that map too. Draw yours first. news.sap.com/2026/10/sap-to-…

When AI fails, the management decision nobody made is the story, not the model. My take: "70% on people, 20% on process and 10% on the technology. The technology is always the easy part." Future Stack Ep. 26 with Paul Lewis: youtube.com/watch?v=w4bIUrjt…

Your head of AI has accountability without authority. You've hired a scapegoat, not a leader. Paul Lewis: "I need a million dollars to do this thing, and I won't do it for 900." Future Stack Ep. 26 with Paul Lewis: youtube.com/watch?v=w4bIUrjt…

OpenAI says its agent will bring you work "done the way you would do it." Great, if you ever wrote that down. Most of us haven't. In my world, that page is the baseline. The agent builds what's on it, with your name on it. Field Notes 03: youtu.be/kjAxHJxRDQg

Copilot didn't leak your data. Your permissions did. You had no rules. Generative AI doesn't fix that. It's a magnifying glass on the problem. Future Stack Ep. 26 with Paul Lewis: youtube.com/watch?v=w4bIUrjt…