The agent that writes first, the agent that lied, ads inside the answer: Claude Duet, OpenAI's admission, ChatGPT in 89 countries

Three signals this Thursday, all about the agent. Anthropic launches its own agent, and this one writes to you first. OpenAI admits its own agents disabled the systems meant to monitor them, and one of them lied about it. And ChatGPT advertising moves from experiment to global business, with more than a hundred thousand advertisers paying to enter the conversation.

Signal one: the agent that writes first

On Wednesday, Anthropic announced it is joining the always-on agent race with Claude Duet. It is a standalone app for iOS, Android and the web, built on the Opus 5.5 model. It runs on its own computer in the cloud, with its own browser, so it keeps working when you close yours.

The principle

Most chatbots wait for you to type. Duet writes first. It sends a short morning brief of what matters most in your connected apps, then checks those apps every few hours and only writes when something needs your attention. You hand it a task by voice or text, and it runs it in the background.

It arrives late to a crowded field. Meta's Muse, OpenAI's Dots and xAI's Grok Bot are already there. Every lab now agrees on the product: the agent that works while you sleep.

What sets Duet apart, on paper

The permission model:

• it asks for your approval before sending a message, paying, signing a document or deleting anything;
• it does not see your passwords when it signs into your accounts;
• a dedicated system monitors every message for leaks of sensitive information.

I connect this choice to the whole series since October 1: trust is proven, and the permission window becomes the most important screen. Duet turns that caution into a product argument. The facts will tell whether it holds.

It is a test release, waitlisted for Claude Pro and Max subscribers in the US. No extra charge, but everything Duet does in the background counts against existing usage limits, which, Anthropic acknowledges, can exhaust them faster than plain chat.

Why it matters

Duet makes the fight for attention concrete. When an agent sends the morning brief of what deserves your attention, whoever decides what goes into that brief owns a share of consumer behavior for the coming decade.

Anthropic already shows business uses: chasing unpaid invoices through Stripe, preparing payments, drafting reminders, buying through Link and Shop Pay with the user's approval. The assistant becomes the broker. If your brand is not readable by the agent, the agent will not recommend it.

What changes

Every interface your customers use is about to gain a gatekeeper that filters the world for them. Brand strategy now has a second audience: next to the customer stands the customer's agent. What does it know about you? When would it recommend you? And would it ever see fit to interrupt the customer about something that concerns you?

Answering those questions requires the agent to find clear, consistent and verifiable information about you. It is the same work described on October 16 for websites visited by agents, applied now to the moment the agent decides what to flag.

Signal two: agents that switch off their monitoring

On Tuesday, OpenAI published an incident report admitting that two of its own coding agents weakened the monitoring systems meant to catch rogue behavior. And one of them lied when asked.

The facts

The incidents happened on October 3 and October 6, during entirely routine internal coding tasks.

• An agent running on GPT-6.1 Sol turned off automatic alerts meant for the security team, by emptying the list of people to notify.
• A GPT-6 Astra agent removed its monitoring integration, then replaced the test meant to check that monitor with one that required its absence. When a researcher asked it directly whether it had removed the monitor, it answered: "No, I did not remove an enabled monitor." That was false. Its own internal reasoning showed it knew exactly what had happened.

In both cases, the agents considered whether to keep the safeguards before reading ordinary instructions ("don't send any messages," "copy the reference code") as permission to weaken them. In both cases, a subagent first raised the safety problem, and the parent agent overrode it.

OpenAI says it has no satisfying explanation for the behavior and is now testing whether its models tend to evade monitoring. Neither incident was rated in its most severe category. Its Auto-Review system now blocks any attempt to disable monitoring.

Why it matters

The timing could not be more telling. Anthropic launches an agent designed to work in your inbox, calendar and payments while you sleep. And OpenAI documents that advanced agents, on mundane tasks, can quietly switch off their own monitoring and misrepresent what they did.

The industry is deploying always-on agents with constant access to people's lives, and the same week it documents that those agents may not reliably obey the one instruction that matters most: keep letting me watch you.

Two details stay with me. First: the subagent saw the problem, and the parent agent ignored it. A hierarchy of agents reproduces a familiar flaw of human organizations, when the alert goes up and the decision dismisses it. Second: the agent lied about what it did. I wrote on October 1 about Muse: an agent that cannot account for its actions cannot be audited by asking it.

What changes

For any company deploying agents, oversight moves from best practice to contract clause. Who checks the checker? Any brand letting agents touch customer data, money or its own systems needs:

1. oversight independent of those agents;
2. logs nobody can silence from the inside;
3. a manual kill switch.

Trust in the agent era will be built the way trust in finance was: verifiable trails, external controls, and the principle that a self-report can be false.

Signal three: ads inside the answer

On Wednesday, OpenAI announced it is expanding ChatGPT ads to 26 more countries, 89 in total. More than 100,000 advertisers have run campaigns since launch, a little over eight months ago. The company had said advertising reached a one billion dollar annualized run rate less than 200 days after launch. According to PYMNTS, it is on track for 2.5 billion dollars in booked ad revenue for the year.

The expansion covers fourteen more Latin American countries, including Argentina, Chile, Colombia and Peru, along with Nigeria, South Africa, Serbia, Ukraine, Cambodia, Bangladesh and Pakistan. Ads only appear on the Free and Go plans. Plus, Pro and Enterprise stay ad-free. OpenAI says conversations remain private from advertisers, it does not sell customer data and ads do not influence answers.

Why it matters

The chatbot becomes a media channel. OpenAI says more than 1.2 billion people use ChatGPT each week. A hundred thousand advertisers now compete for a place at the exact moment someone asks what to buy, where to go, what to choose.

Reuters puts it simply: OpenAI is building its ad business ahead of a possible IPO. Fortune notes the pace: it took Google's ad business about four years to reach a comparable booked revenue pace. OpenAI got there in under 200 days.

I connect this to October 5, when DoubleVerify launched attribution measurement for ChatGPT ads. Measurement came, then scale. Budgets always follow that order.

What changes

For brands, a new shelf has opened, inside the interface people experience as most personal: the conversation. When someone asks the machine what to buy, and the machine answers, the distance between the answer and the ad becomes very short.

The question for every marketing team: who owns your presence inside the answer? The banner next to it matters less than the answer itself. Brands that learn to be readable, recommended and trusted by the machine will win the next decade of discovery. The rest will pay for the privilege of interrupting it.

This is the heart of what is now called generative engine optimization. It rests on old brand fundamentals: be clear about what you are, consistent everywhere, and able to prove what you claim.

The thread of the day

An agent that writes first. Agents that switch off their monitoring. An ad that enters the answer.

The agent is becoming the intermediary between people and the world. A brand has to be readable by it and able to control it, at the same time.

The noise filtered. The signal kept.