Sponsored by
Hey,
A Washington Post columnist sprinted through O'Hare, reached his gate on time, and was turned away. His seat had been given to someone else. Not by a person, and not because he was late.
⚠ THE STORY OF THE WEEK
American Airlines runs a system called AURA, for Automated Reaccommodation. It predicts which passengers will miss a connection, rebooks them onto a later flight, and releases the seat they were holding.
The prediction does not have to be correct for the seat to be gone.
Marc Thiessen posted about it on August 10. Media consultant Beverly Hallberg replied that the same thing had happened to her the day before. Neither was asked. An internal memo described the logic as finding "discovered inventory" from passengers "certain to misconnect," and acknowledged that affected flights may temporarily look overbooked.
Two details make this more than an airline story. Gate agents reportedly cannot reverse what the system decides, so the human in front of you has less authority than the model. And AURA has been live since 2023, expanding to six hubs this year. It has been doing this for three summers. What changed in August is who noticed.
This is the shape of the thing everyone is now shipping: an agent with real authority, optimising a metric that is not your metric, with no override path. Keep that in mind for the rest of the issue.
01 · Everyone else is building the same thing
🏭 Toyota is running 50+ production agents. North America cut its AI delivery cycle from roughly six months to four days using Deep Agents and LangSmith, and is tracking ROI per agent. This is the competent version of the same pattern, and it is why the airline version will not be the last one you hear about.
⚖️ Regulators arrived faster than usual. Google published its Agent Payments Protocol (AP2), NIST is working on agent identity and permissions, and the AI AGENT Act (S.5051) is moving. The common thread is verifiable, task-bounded authorization: a signed record of what an agent was allowed to do, travelling with each request.
🔒 Google Cloud's advice, in one line: do not bolt agents onto legacy access and logging. Short-lived permissions, task-level traces, and a human escalation point, before you scale anything autonomous.
If you are handing any tool the ability to act on your behalf, the question stopped being "is it smart enough." It is "what is it allowed to do, and how do I take that back."
02 · Today: o3 leaves ChatGPT
⏳ TODAY, AUGUST 26
o3 retires from ChatGPT, closing the 90-day sunset that began with the May 28 notice. The standalone DALL·E GPT is going as well.
Both are ChatGPT-surface changes only. The o3 notice says explicitly that the API is unaffected, and the DALL·E retirement removes the standalone GPT, not image generation, which continues in ChatGPT Images.
If a saved prompt, custom GPT, or team workflow is pinned to either, it breaks today rather than next week.
Customer agents you can test, version, and trust in production.
Most teams treat AI agents like a black box. They wire together a workflow tool, a speech vendor, and a transcription API, then hope the result behaves. When it doesn't, there's no way to know why, and no safe way to change it.
ElevenAgents treats customer agents like software. Built on the voice models the market already builds on, it runs voice, chat, email, transcription, and reasoning in one pipeline, so a customer can start in chat, escalate to a call, and follow up over email without losing context. Voice responses come back in under 400 milliseconds.
Then you get the controls stitched-together stacks can't offer. A/B test prompts and personas with Experiments. Enforce behavior with Guardrails. Version every change, ship improvements, roll back mistakes. Plug in any LLM and ground answers in your knowledge base.
Launch in minutes, improve every week. Transparent, flat pricing at $0.08 per minute.
The ToolChase site has no ads, no pop-ups, and no cookie wall. Two labeled sponsors here are what pay for that, which is also why our scores stay unbuyable. A click costs you nothing and keeps the whole thing running. Want a slot? Here's how it works.
03 · The Stack Builder is now everywhere
✦ NOW LIVE SITEWIDE
Pick your job. Answer five questions. Get a ranked stack.
We shipped it on the homepage last week. It now runs across the guides too, so it meets you wherever you are researching instead of making you go back to the front page. 64 jobs across 9 verticals, scored against 1,900+ hand-verified feature grades, with the reasoning shown for every tool.
Free. No account. Deterministic, and blind to affiliate status.
The connection to the rest of this issue is not accidental. Every system in the story above decides for you and hides the reasoning. The whole design brief here was the opposite: you pick the job, you see why each tool ranked where it did, and you can disagree with it. A recommendation you cannot inspect is just an instruction.
Reply and tell me what it got wrong. The bad recommendations are the useful ones.
04 · Also worth knowing
⚡ Nvidia's Groq 3 LPX entered full production. The inference accelerator built out of its $20B Groq acqui-hire, up to 256 per rack in the Vera Rubin platform, with Nebius as first cloud customer. Nvidia also disclosed the Vera CPU at Hot Chips: 88 cores across six chiplets, tuned for orchestration and tool-calling rather than raw core count. Translation: the hardware is now being designed specifically for agents.
📰 Thomson Reuters launched its own frontier model. Another domain incumbent deciding it would rather own the model than rent one. Expect more of this from companies whose moat is proprietary data.
Full running log at AI Tools News 2026.
Blu Dot surpasses 2,000% ROAS with self-serve CTV ads
Blu Dot used Roku Ads Manager to drive incredible results for its furniture sales event. Its strategy hinged on custom audiences and retargeting, where intent was strongest.
“Roku has been a top performer,” said Blu Dot’s Claire Folkestad. “We have seen…CPMs lower than any other CTV partner we've worked with.”
05 · Prompt of the week: the permission audit
The airline story is only interesting because most of us have quietly granted similar authority to tools we use daily. Worth ten minutes to find out what you have actually agreed to.
You're a risk analyst. Here are the AI tools and services I have connected to my accounts, and what each can reach: [LIST: tool → what it connects to].
1. For each, what actions can it take without asking me first? Distinguish read-only from anything that writes, sends, books, buys, or changes state.
2. Where an action is automatic, what is the reversal path, and how long do I have?
3. Which of these could take an action that costs me money or time, based on a prediction rather than a fact?
4. Rank them by blast radius: most damage possible with the least oversight.
Flag clearly where you are uncertain about a specific tool's behaviour rather than guessing, and tell me what to check in its settings to confirm.
Why it works: point 3 is the AURA question. Any system that acts on a prediction will sometimes act on a wrong one, and the cost of that lands on you, not the system. Knowing which of your tools work that way is the whole exercise.
06 · Three guides worth your time
🤖 Best AI Coding Agents 2026
The category where "acts without asking" is a daily decision. Which agents let you gate what they merge.
🎨 Best AI Image Generators 2026
With the DALL·E GPT retiring today, a good moment to know what else is worth your workflow.
🆓 Best Free AI Tools
Genuinely free, not free until the one feature you need. Our most-shared list.
THE TAKEAWAY
An agent that acts on a prediction will sometimes act on a wrong one. The question is never how smart it is. It is what it can do before you get a say.
Reply and tell me: what is the most autonomous thing you have let an AI tool do on your behalf, and did it go well? I read every one.
See you next Wednesday,
Emre
Founder, ToolChase
Building an AI tool? Submit it for a free editorial review. Read The State of AI Tools 2026: 724 tools analyzed, only 6% truly free.