#agents - natmc

Friday, October 9, 20261 item

X DEBRIEF

FRIDAY, OCTOBER 9 - 2 posts · tagged: #ai #agents #insurance

DID THE PROMPT ACTUALLY SAVE MONEY?

David Goldberg posted a prompt for having an AI agent haggle over internet, phone and insurance bills. I asked the question that comes after the prompt: "Good prompt. Have you realized any savings yourself from it?"

His answer: "Yup. Got insurance bill knocked down."

The proposed job went past looking up cheaper plans. It asked the agent to find what new customers pay, contact each provider by chat or email, and ask for the same rate. It also told the agent to keep escalating, including threatening to switch or cancel.

There was an explicit stop in that instruction: no cancellation without approval. The prompt asked to hear from the agent only if the provider needed the customer's voice.

Goldberg reported a lower insurance bill, but his reply did not give the amount saved, the insurer, or the steps that got the reduction. The result was narrower than the prompt: one reported insurance saving, not a breakdown across all three bills.

Monday, October 5, 20261 item

X DEBRIEF

MONDAY, OCTOBER 5 - 3 posts · tagged: #ai #dealyo #agents

THE INTRODUCTION WAS ALREADY IN THE NOTES

Jesse Middleton asked what happens if the major AI models stop improving. His answer: we'd still have years of work figuring out what to build with them.

His example was an introduction. After a founder call, Muse read his Granola notes, noticed he had offered to connect two people, and drafted the introduction. He hadn't separately asked it to write the email. The commitment was already in the meeting notes.

That's a useful product question: can the assistant pick up work you already agreed to do? It has to find the commitment, understand enough context to draft it, and earn enough trust for you to keep giving it access. Jesse was talking about making that happen reliably, rather than waiting for another model release.

He invited people building with today's AI to show him what they were working on. I replied: "You might find what we’re working on at https://dealyo.co interesting. We should do a proper catch up!"

Friday, October 2, 20261 item

X DEBRIEF

FRIDAY, OCTOBER 2 - 3 posts · tagged: #ai #agents #shopping

PROVE YOU'RE NOT THE ROBOT I HIRED

"Prove you're not a robot," said Target, to the robot I hired to buy a $2 frame.

The robot was the point. I wanted it to do the shopping. Target wanted it to prove it wasn't the thing I had hired. "Really wish that trad Ecom was more friendly to ai bots."

Clay Wolcott replied: "It ruins the fun"

Later, Rob Leathern described a 1,200-site web search job with Muse. It took about six hours and brought him to 40% of his weekly token limit. I asked: "Interesting - can you share what kind of sites you are looking at and why?"

He couldn't share the sites, but said it was "to get a single small piece of data" and "a successful experiment".

Rob had a six-hour search job to report. I had a robot being asked not to be a robot. His agent was looking through 1,200 sites; mine was trying to buy a $2 frame. The frame wasn't an ambitious test of what AI could do. It was a small shopping errand, which is why the robot check was such a ridiculous interruption.

Saturday, September 12, 20261 item

X DEBRIEF

SATURDAY, SEPTEMBER 12 - 5 posts · tagged: #agents #cfoai #tesla

DIRECTIONS FOR AI AGENTS, LINKED FROM THE HOME PAGE

At 11:52am I called out one thing I like about blader's cfo.ai (re)launch: "they have directions for AI Agents linked to from their home page" (x.com). His launch post says runway is now cfo.ai, that you can hire @arithecfo and he can start today, and that you can test-drive him by replying to the thread with anything ridiculous you want modeled - he comes back in 30 minutes or less with a fully wired financial model. My take: "I do think the Agent Led Sales will be a big thing going forward. Lean in today."

At 7pm I posted the general version: "When an agent hits a wall, asks for permission to buy access, and keeps going, that is owning the outcome. The app layer will belong to systems of execution that actually close the loop" (x.com).

In between, Tesla posted "Go for launch" - a taillight photo stamped 10.01 - and Jason Baptiste quote-posted it: "Finally. It's time." Me: "Holy sh*t!" (x.com). Jason offered: "I will take you for a ride." My question: "Founder edition?!" (x.com).

Sunday, September 6, 20261 item

X DEBRIEF

SUNDAY, SEPTEMBER 6 - 5 posts · tagged: #ai #agents #instinct

THE VERB CHANGES HANDS

A quiet holiday Sunday produced the week's shortest thesis. Mid-afternoon, in one line: we are moving from software that helps us know what to do to software that actually helps us do it - sources, then answers, then outcomes (x.com). Just after midnight came the working proof, offered as casually as a shrug: using an agent more and more to say "pls do this.." - and often it can (x.com).

The progression is worth memorizing because each step hands over more of the verb. Sources: the machine finds, you read. Answers: the machine decides what the finding means, you choose. Outcomes: the machine does the thing, and you find out after. Nine words - sources, answers, outcomes - and each one moves the human one seat further from the controls. The midnight reply is what the third step feels like from inside: not a demo, not a roadmap, just a person saying please and being surprised, still, that it works.

The counterweight arrived an hour before that reply, as a joke. A video bit about a Cybercab driving a passenger straight to jail because the onboard model found a warrant, and the response: I know it is a joke, for now, but it does not seem far fetched (x.com). The parenthetical is the whole column. For now. When software only knows things, its mistakes are wrong answers. When software does things, its mistakes have doors and wheels. The distance between "pls do this" and "drove me straight to jail" is not intelligence. It is permission - and permission is the part nobody has written the manual for. If the progression holds, the next version of that manual gets written by whoever learns to say please carefully.

Friday, September 4, 20261 item

X DEBRIEF

FRIDAY, SEPTEMBER 4 - 18 posts · tagged: #ai #agents #commerce

WHEN SOFTWARE BECOMES A CUSTOMER

Two things happened within three hours on Friday afternoon. First, a field report: an agent hit a limit on a lead scraper, found the paid tier, and asked to open an account with a credit card (x.com). Then, reading what season three of a Netflix show cost, the prediction that someone would fire up a video model over the weekend and try to make it for one-hundredth the price (x.com). Neither is a headline. Together they mark one change: the software stopped only doing the work and started taking part in the market.

Sit with the number. A hundredth is not a discount; it is a different economy. When the cost of making something falls a hundredfold, the budget stops being the barrier, and the scarce thing becomes the judgment to know what is worth making. And sit with the progression in the field report. The agent hit a limit. It found the paid tier. It asked for the card. Each step is small. Together they are the difference between a tool and a customer - software that notices a price, decides it is worth paying, and reaches for the wallet.

The tension is that the public argument is still stuck on the old question. The same afternoon brought a sarcastic jab at the latest study: so AI creates jobs? (x.com). Count the jobs created, count the jobs taken - the scoreboard argument assumes the players are still people and companies. But the afternoon's evidence says the players are changing. Agents now hold cards, pick tiers, and bid against studios. The question is not whether AI creates jobs. It is what a market does when the newest customers are not human. If this continues, prices will start clearing between machines, and the next negotiation will not involve a person at all.

Wednesday, September 2, 20261 item

X DEBRIEF

WEDNESDAY, SEPTEMBER 2 - 28 posts · tagged: #ai #agents #instinct

THE TIE AT THE TOP

The day opened on the model race: the SOTA models are so neck and neck that it is surprising none has pulled ahead (x.com). By afternoon the point had sharpened. Gemini 3.8 Flash is basically frontier-class on a surprising number of tasks at a fraction of the price (x.com). The answer to a direct question about Kimi drew the same map: all the models are competing fiercely, there is no clear winner, and specialization and price will decide it (x.com). The race at the top has ended in a tie.

A tie at the model layer moves the value to the work layer. The midday thesis: the biggest shift in AI is from charging for software to charging for outcomes, and replacing $30-a-seat workflow tools with systems that directly execute the transaction is where the trillion-dollar companies get built (x.com). The rest of the day read like evidence. Aslan AI raised $20.8M to run undercover agents in criminal forums for the FBI (x.com). fal cut the price of video generation again (x.com). One 12-person company is spending $2,000 to $4,000 a month on an SMS assistant just to automate basic ops (x.com). And the working proof: an agent planned a meetup end to end - emailed eight NYC venues, caught its own bounced emails, fielded the replies, priced three private rooms, and locked a back room at Linen Hall for September 15. The first NYC Instinct event is on the calendar and the RSVP is open (x.com). One reply added the missing word to the outcomes thesis: speed (x.com).

The last post of the night put a ceiling on the whole picture: 80% of OpenAI and Anthropic's enterprise revenue comes from 1% of customers (x.com). The easy read is that the labs are capped. The better one is to imagine the other 99% growing into how the pioneers already use these tools. The lesson of the day: the scoreboard is a tie, so stop watching it. Count what ships.

Tuesday, September 1, 20261 item

X DEBRIEF

TUESDAY, SEPTEMBER 1 - 7 posts · tagged: #ai #poker #knicks #agents

WHEN +EV ISN'T THE WHOLE QUESTION

September's AI release calendar, laid out in one quoted list - Fable 5.1, GPT 6 Astra, Grok 4.7, DeepSeek V5, Gemini 3.8 flash and more inside a single month - drew the day's first reaction: spelled out like that, the pace of change is remarkable (x.com). Hours later Every's week-long Fable 5.1 test landed - "the strongest coding model we've used" - and got a repost plus a short verdict: "This review makes me want to use Fable 5.1. Check it out." (x.com).

The day's most engaged post was a poker hand. Sixteen players left, A8s in early position with ten big blinds and the blinds about to jump - shove, button shows queens, out after three days of grinding. The question isn't the shove's chip EV, which was probably fine; it's whether a decision can be right in isolation and wrong once you weigh what happens if it goes badly - startups included (x.com). One reply pushed back directly: "pretty clearly the right call" (x.com).

Elsewhere: a webhook that returned HTTP 200 while a Postgres worker silently dropped the payload over an unescaped unicode character in a property address - "My grandmother would think I am speaking a foreign language" (x.com). An inbox agent that finds unopened newsletters, adds checkboxes, and unsubscribes on command: "works like a charm!" (x.com). A clear-grape Gatorade review from a glass-bottle lifer who wanted to hate it and kinda digs it (x.com). And a note that Jalen Brunson turned 30 - if he gets another ring, "they better just give him the statue" (x.com).

Monday, August 31, 20261 item

X DEBRIEF

MONDAY, AUGUST 31 - 2 posts · tagged: #venture-capital #agents

THE COST OF SELLING TOO EARLY

Doug Leone's admission is the kind of number venture capital usually keeps quiet: Sequoia owned 10% of Google, 20% of Nvidia, and early Apple, and sold all of them. The post frames it with Fred Wilson's long-running argument that fund structures force firms to sell their best assets too early. Treating software like a 10-year fund cycle instead of a multi-decade compounding machine left hundreds of billions on the table. The structure, not the picking, was the mistake.

The late post is a public experiment log: an agent handed an open-ended errand - find 5-15 bars near Union Square and Flatiron, dig up private-events contacts, organize a 20-40 person meetup with a cash bar and private room on a $250-300 minimum spend, pick an open Tuesday in mid-September, then stand up a Luma invite and find places to submit it (link). Results promised in the morning. The instruction set reads like a brief to a competent assistant, which is exactly the capability on trial.

Sunday, August 30, 20261 item

X DEBRIEF

SUNDAY, AUGUST 30 - 4 posts · tagged: #ai #knicks #nyc-eats #agents

WHAT HAPPENS WHEN THE WAIT GOES AWAY

Two of Sunday's posts circle the same question. Taalas is generating roughly 14,000 tokens per second, about 100x ChatGPT's pace, and the post quoting the demo asks the follow-up: what happens when waiting goes away? The evening's first post answers it in advance: demand for intelligence grows 50x in the next few years. If both hold, faster supply doesn't relieve demand - it creates it.

The rest of the day: Kevin Durant spent the summer publicly crediting the Knicks' Finals comeback with breaking his brain, an unusual amount of praise from that corner. And a documented home attempt at L'Artusi's mushroom dish, rated pretty good.