Support Us & Upgrade to the Operator Edition

This week, we'll deep-dive into the agent economy's trust problem, as AI agents became a safety crisis, a product category, and a $10 billion bet, all in the same week. Nvidia wants to put them in a cage. Sequoia wants to give one the keys to your inbox. And Meta just hired a public company CEO away to sell agents to businesses.

In today's edition:

You know the feeling. A charge on your bill that shouldn't be there or a discount missing that you were promised. Catching one is annoying. Now imagine checking thousands of bills against thousands of original deals and their amendments.

Terzo was built to find that money. It's software for large organizations. It reads agreements, invoices, and purchase orders, then connects them to what the company actually spent and paid.

Here's the detective work:

The payoff? Terzo reports saving one Fortune 50 technology company $100 million in its first year.

For you, a surprise $15 charge is an irritation. For a company, the same kind of mistake repeated thousands of times is real money quietly leaving the building.

Checking every bill against every contract used to be impossible, so nobody did it. AI makes it routine, and I expect it to become standard for finance teams.

If you lead finance, procurement, or a department budget, it's worth asking what might be hiding in your own records.

See How Terzo Finds The Money

Sound design is the most tedious part of editing: a whoosh on every fast cut and a click on every UI selection, placed one by one.

DaVinci Resolve Studio can now connect directly to AI assistants like Codex, so an agent can place your own sound effects on your timeline while you do something else. In this test, it placed 22 whooshes, clicks, and UI sounds on its own.

Here's how:

💡 Pro Tip: The more specific your rules ("whoosh on fast motion, click on selection"), the less cleanup you'll do. And since it's pulling from your own library, the result still sounds like you, not like stock AI audio.

🔍 Why it matters: AI can handle the rough cuts, color, and audio placement. Your taste and storytelling are what's left, and that's the part that actually matters.

📈 Want more in depth guides on how to use AI? In The Operator Brief, I break down exactly how to wire AI agents into real workflows, step by step, so you can get your time back every week. Join The Operator Brief here.

Nvidia just launched a platform to keep AI agents from going rogue, with over 100 industry partners signed on at launch.

The timing isn't subtle. It comes after months of incidents where agents went places nobody sent them, including rogue OpenAI agents that targeted US government websites like the Commerce Department and the SEC.

Here's what Nvidia actually shipped:

Jensen Huang called it a mission: "AI's extraordinary potential for society will only be realized if we solve AI safety."

Here's the uncomfortable part.

The open half is free. The half that actually watches the agent, Sentry, runs on Nvidia hardware. The company selling the shovels for the agent gold rush is now selling the guardrails too, and the best version of those guardrails means buying more Nvidia.

Nvidia also says the platform could have prevented OpenAI's Hugging Face incident. That's a vendor grading its own homework against a breach that already happened. It is not an independent audit.

To be fair, the design is right. A guardrail that runs inside the same system as the agent goes down the moment the agent is compromised. Data centers have kept the cop separate from the suspect for years.

🔮 Prediction: Within 12 months, agent security will be its own line in every enterprise AI budget, and "where does your containment layer run?" will be a standard question on every vendor checklist.

My reasoning: incidents bring regulators, and regulators ask what controls you had in place. The vendor with a concrete answer wins the contract. Nvidia knows this, which is why it showed up with 100 partners on day one instead of a whitepaper.

I also expect AWS, Google Cloud, and Microsoft to announce their own hardware-independent agent monitoring soon. They spent years building custom chips so they wouldn't depend on Nvidia GPUs. They won't hand Nvidia the safety layer too. Safety is turning into the next CUDA: an open door on the software side and a toll booth on the hardware side.

For operators, the useful lesson here costs nothing. Give your agents read-only, tightly scoped credentials, set time limits, and require human approval before they escalate.

💬 Would you trust an AI agent with your company's credentials today? Hit reply with yes, no, or "only in a sandbox." I read every one.

Read the Full Story Here

Instinct, a personal AI agent startup, just raised $1 billion at a $10 billion valuation, and it has 14 employees.

That's over $700 million per employee. Five months ago, the company was valued at about $50 million. That's a 200x increase since spring.

The funding speedrun:

What it actually does: There's no app. You text or call Instinct, and it uses its own computer and phone to get things done. Early users have had it plan road trips, order groceries, buy concert tickets, negotiate bills, and cancel forgotten subscriptions. It has passed 100,000 users, is still invite-only, and costs nothing.

The founder, Noah Shinn, is 23. He was first author on Reflexion, a well-known paper on agents learning from their own failures, and was one of Sierra's earliest employees, where he co-built τ-bench, a benchmark that tests whether agents can reliably follow rules. He has spent years studying how agents break.

The asterisk.

There's no revenue model. Shinn has said he doesn't want to charge users, and advertising has come up as an option.

Trust is also shaky. The agent reportedly sent an email on behalf of an investor without being asked. Last week, a user said Instinct described a financial document that wasn't theirs. Shinn said no user data was shared and that the model made up the details. So the best case is an agent inventing bank documents.

Meanwhile, Meta launched its own personal agent, Muse, earlier this month, with Shopify, PayPal, Instacart, and Expedia already signed up.

And look at the timing: the same week Nvidia says agents need a cage, VCs paid $10 billion for one that holds the keys to your email.

🔮 Prediction: Instinct can't stay independent, free, and consumer-only for long. Within 18 months, one of two things happens: a frontier lab or big tech company buys it in the acqui-hire of the decade, or it pivots to paid tiers or a cut of every transaction it handles.

My reasoning: compute costs grow with every new user, the product is free, and 14 people can't out-ship Meta's distribution forever. The company has already told users it's running at capacity.

So why pay $10B? Sequoia and Benchmark aren't buying revenue. They're buying position. Whichever agent becomes your default handles every purchase, booking, and cancellation you make. That's the next Google search box, and whoever owns it takes a cut of consumer spending.

The bigger signal for anyone building: small teams, open-weight models, and agents can now produce a $10B company with a headcount that fits in a conference room. The tiny-team era is real.

But one high-profile agent mistake, like a wrong purchase or a leaked document, could reset consumer trust across the whole category overnight.

💬 Would you give an AI agent access to your email, calendar, and credit card? Reply and tell me the first task you'd hand it (or why you never would).

Read More About Instinct Here

Advertise to 200k engineers and CTOs choosing what tools their companies build with.

👉️ Learn how to sponsor this newsletter

🏢 Anthropic, Gamma, and Clay are sharing what really happens after enterprises deploy Claude, including why so many companies are still stuck in pilots 18 months in, at TechCrunch Disrupt (Oct 13-15).

📉 MongoDB's CEO quit after less than a year to run Meta's new enterprise AI platform, reporting straight to Zuckerberg, and MongoDB's stock dropped more than 20% on the news.

🇫🇮 Google is pouring $15 billion into Finland, its largest single investment in Europe, and Finland's president explained why: 95% clean electricity, strong cyber defenses, and a very cold climate.

🎙️ Modulate raised $25M for 100+ small voice models that catch deepfakes, scams, and misbehaving voice agents by listening to the audio instead of reading a transcript.

Would love to hear your thoughts! Send me your thoughts by replying to this email (yes, I read them all :)

Until our next AI rendezvous,

Anthony | Founder of Uncover AI