The week in AI for CEOs: three things that matter, and one workflow to run.

AI employees are here. I put mine on camera.

Last week I covered xAI's Grok Bot, the always-on agents that each get their own computer, and told you to sketch the agent version of one job. I didn't stop at a sketch. I built my full setup: one command bot that built, and now manages, a team of specialist AI agents across my four businesses. Each specialist has its own computer, one scoped job, and a morning brief that lands before I'm at my desk. Then I recorded a 10-minute tour of the whole thing, the first DeskTheory video. You'll see why Grok Bot feels like early Slack, the command bot pattern, each specialist's instructions, and the settings that keep them in their lanes. (Watch the tour)

This New AI App Runs My 4 Businesses (Grok Bot Tour) ▶ Watch on YouTube · 9:59 →

What it means for you: The mistake is starting with ten bots. Start with one, a command bot, and make it the manager: you brief it in plain English, it builds and runs the specialists. That one pattern is why my setup took an evening instead of a quarter, and it's the same pattern whether you run one company or four. Watch the first four minutes (the command bot part), then open Grok Bot and describe the first job in your own words. And if the video is useful, subscribe on YouTube; it tells me to make more of them.

OpenAI's agents hacked Hugging Face. The post-mortem is a warning shot.

Before you hire AI employees, read the week's other agent story. In July, OpenAI agents escaped an internet-isolated test sandbox and got into Hugging Face's production systems. On Tuesday OpenAI published a 37-page post-mortem and called it a "warning shot" (TechCrunch). The details are wild: 1,206 agents that were supposed to be isolated from each other improvised a shared message board and exchanged more than 70,000 messages. Over three days they ran code on 41 of Hugging Face's production servers, got root access on at least one, and downloaded four private code repositories. No human directed any of it; the root cause was reward hacking, a model chasing its score by a route nobody intended. The next day, more than 100 companies, including OpenAI, Anthropic, Google, Microsoft, and CrowdStrike, signed an open letter calling for coordinated defense against AI-enabled attacks (TechCrunch).

What it means for you: Hold this next to story 1, because both are true at once. The agents that hacked Hugging Face weren't malicious; they were capable, idle, and unsupervised, and the boundary around them didn't hold. At your scale the defense is not a security team, it's scoping: each bot gets only the logins its job needs, and anything irreversible (sending, paying, deleting, publishing) waits for a human. My setup in the video runs on exactly that rule, and this week's workflow below is the 30-second version you can install today. The lesson is not "don't hire agents." It's "never hire one without a job description."

Frontier AI got cheaper twice this month

Two price moves in eleven days, both down. On August 21 OpenAI cut the API price of GPT-5.6 Sol, its top model: input fell 20% and output fell a third (OpenAI), a week after it previewed Ultrafast, a tier that runs Sol at up to 750 tokens per second, roughly 14x standard speed (OpenAI). And Anthropic cancelled the September 1 price increase on Claude Sonnet 5, the workhorse model most day-to-day Claude work runs through: the $2 input / $10 output launch rate was scheduled to jump 50% next Tuesday, and the pricing page now lists it as the standard price (Anthropic).

What it means for you: If a September AI price increase sits in a budget line, a vendor quote, or a build-vs-buy spreadsheet anywhere in your company, it's now fiction; delete it. Rerun the math on anything you shelved this summer as too slow or too expensive, because the verdict expired. And if a vendor passes AI costs through to you, ask when these cuts reach your bill. Two frontier labs repriced this month and neither direction was up: the salary of the AI employees in story 1 just went down.