Jason: AI代理改变获客、高管挖角与出售决策的9条实战教训

My cut of the episode where Dev Ittycheria, one week back as CEO of MongoDB, told me I was wrong about the biggest story of the week.

Thursday’s recap covered all four of us. This is my side of it: ten decisions I’d make sooner after an hour with Harry, Rory, and Dev.

#1. A Board Observer Became the Competitor’s CRO in Two Hours, and Dev Told Me I Was Wrong

Factory’s CEO posted that he had fired adviser Chris Degnan over talks with Cognition, and two hours later Degnan announced he was Cognition’s chief revenue officer. Degnan says he resigned and shared nothing.

My view on the show: 95% of CROs would say this is how they found their last job. He was an adviser and observer, with no board seat and no employment contract. A lot of CROs treat competition as a game and are friends with their counterparts across the street.

Dev disagreed, and his argument deserves a fair statement. He has never seen a CRO go to a direct competitor. A CRO recruits people on the vision and on what it will mean for their families, and leaving for the rival tells every one of them they were sold a bill of goods. His question to any departing executive: did it have to be a direct competitor.

I still think the culture has changed permanently. Half the frontier-lab talent rotates between labs, the compensation is different, and people move faster. It can bother you. The world changed anyway.

My learning → Assume your CRO, your best engineer, and your most plugged-in adviser each get an offer from your direct competitor this year. Dev got his version of that call the week before he came on the show.

#2. Execs Used to Take Two People With Them. I’m Now Seeing Eight to Ten in the First Week.

For most of my career there was a rule of two. You left, you took one person, your old CEO got angry. You asked permission for the second and there was friction. Cross two and the relationship was broken.

That rule is gone. I’m seeing people I respect, people with high ethics, take eight to ten in week one. Dev has lived the carpet-bombing version and said it came with stern words. Harry asked him which is worse: an exec who goes to a direct competitor and recruits no one, or one who goes somewhere unrelated and takes five of your best. Dev said it depends on whether the competitor is his mortal enemy.

My learning → The day a senior leader resigns, write down the eight to ten people they hired and talk to each one that week. Under the old rule you only had to worry about two.

#3. Factory’s Adviser Sat In on Board-Level Discussions for About a Year

Degnan had advised Factory for about a year. Dev drew the line I’d use. An adviser you call to vet a hire in Europe is a light relationship, and if that person joins a competitor, you’re unhappy and you move on. An adviser who sees your product roadmap, board plans, and win-loss data is inside the tent.

Rory’s point was that Factory had an implied understanding and nothing explicit. Nobody wrote down what happens when a competitor calls.

My learning → Two tiers. The adviser who makes intros gets a phone call. The adviser who sees the board deck gets one written sentence: if a competitor approaches you, you tell me that day and you stop attending until it’s resolved.

Screenshot

#4. I Asked an Agent Where to Host an App, Live on the Show. It Picked Render.

Vercel says agent-triggered deployments went from under 3% to more than 50%. Guillermo Rauch didn’t build Vercel for that. He had the right infrastructure when agents showed up, and now agents choose it.

I run 20-plus agents and build 8 to 10 hours a day. My agents have opinions and they are hard to argue with. During the episode I asked one what to use for hosting an app I’m building. It ranked Render first and labeled it “my pick.” Railway was second, Fly.io third. Vercel was fifth, marked not suitable for this use case. I would have picked Vercel myself.

My learning → Render got a qualified lead from me with no rep, no ad, and no search result. I barely knew the company’s name before the agent said it.

#5. My First Agentic Purchase Was Resend, Because I Couldn’t Get SendGrid to Work

Last year I was setting up agents and kept getting pointed to SendGrid. The free tier had been deprecated, I couldn’t sort out the key, and it kept breaking. I asked my agent what to use. It said Resend. By the numbers I had on the pod, Resend’s MCP calls went from 106,000 in April to 3 million in September.

Dev made the point that should worry every startup. The incumbent has 10 to 15 years of content for an agent to read, and you have very little. Your docs and your API have to make you a first-class choice for a buyer that never sees your homepage.

My learning → Find 10 people you trust who have agents in production. Every two weeks, ask each to have their agent recommend a vendor in your category, and write down where you rank. I don’t know of a product that does this yet.

#6. Listen Labs Sold for a Reported $2 Billion at Three Years Old. I’d Tell a Founder to Take It.

Salesforce is buying three-year-old Listen Labs for a reported $2 billion. The company walked away from a signed $125 million Series C at $1.5 billion to do it, on about $30 million of annualized revenue.

From the Salesforce side I don’t see how it moves the needle at their scale. From the founder side it’s an easy call. Company-building comes in five-year chunks and each one takes it out of you. I sat in a meeting this year with an offer near this price, and every VC at the table said don’t sell.

The test is whether you’re building something generational. If you aren’t, the company has a terminal value and you won’t beat the net present value of a fair offer. Dev is back in the CEO seat 12 years after he first took it. This stuff isn’t easy.

My learning → Answer one question before the board meeting: is this generational. If the honest answer is no and the offer is $2 billion inside five years, take it.

#7. Every Published Eval Now Needs Three Asterisks and Four Daggers

Reflection announced Beam, a US open-weight model it says matches Z.ai’s GLM-5.2 on reasoning benchmarks with three to four times less inference compute. Those claims haven’t been independently verified.

The demand is real. At Dreamforce, nobody I talked to wanted to run a Chinese model, and the ones who did felt pushed there by cost. I still don’t trust a launch-day eval. I checked OpenRouter during the show to try Beam and it wasn’t listed.

Switching models is easier than switching databases, and I’ve recently done the database. It is still work. You requalify prompts and redo workflows.

My learning → Before moving a workload to a cheaper model, run it on one real workflow where a wrong answer costs you something. Mine is an agent telling a customer an item is in stock when it isn’t.

#8. ElevenLabs Doubled to $22 Billion, and Voice Is Where I’d Still Pay Full Price

ElevenLabs hit $22 billion in a $300 million employee tender, double its $11 billion valuation from February.

I’ve said cost would drive substitution in models this year, and Harry called me on it. Voice is my exception. I can tolerate a little slack in reasoning on a non-critical workflow. When a flower shop’s agent picks up the phone, it has to answer in seconds and get the order right, or the product is worthless. I was surprised by the deal sizes: modest companies paying hundreds of thousands of dollars a year and telling us on reference calls it’s a good deal.

My learning → Split your workloads into two piles. Anything a customer hears or sees in real time gets the best model you can buy. Everything else gets routed to the cheapest one that passes your own test.

#9. Nvidia Calls Its $20 Billion Groq Deal a License. Two Former Engineers Call It a Merger.