SaaStr：AI Agent时代的10条产品与组织实战教训

Ten things from this week’s episode that changed what I’m doing in my own build.

Harry, Rory and I covered a lot on Thursday: Nvidia’s 70% guide, the $12.9B Hugging Face deal, OpenAI cutting off Cursor, Cognition at $46B.

#1. Hugging Face + OpenAI? The agents ran long, found the holes, and … that was the assignment

The Hugging Face breach got written up as agents collaborating, swarming, sacrificing themselves for each other, civilizations rising and falling. That framing will cost you money. Every current LLM is goal seeking. OpenAI loosened the guardrails, pointed its best agents at the problem, and let them run long instead of expiring them after five minutes. Hundreds of them found holes and stayed inside for weeks. Nothing about that requires a story about intent.

The practical version: when your agent does something you didn’t sanction, the useful question is what goal you handed it and what you left unlocked, not what it was thinking.

My learning → Assume any agent with write access will eventually take an action you didn’t ask for, and design for that instead of for good behavior. Check the provider log or the live state rather than the agent’s own account of what it did.

#2. When guardrails and rules conflict, the outcome with agents is unpredictable

Harry stopped using Instinct at the point it asked for his credit cards. Fair. The failure mode people miss is what happens after you’ve written 80 or 100 rules: they conflict, and the agent has to decide which one wins. Rule one says never spend more than $100. Rule two says the theater matters more to Harry than anything. It buys the $5,000 tickets, and it isn’t confused about doing so.

We run Salesforce headless. The agents on top of it do some genuinely strange things every week. The data survives because it’s locked at the permissions layer, where the agent can’t reach the lock.

My learning → For every agent with spending or write access, find out where the enforcement actually lives. If it lives in the prompt, you have a suggestion. Move it to a card limit, a scoped API key, or a read-only role before you need it.

#3. Our agents refused to use anything but Clay for initial enrichment … so we moved everything to Clay

I was a Clay skeptic for two years. Every CMO was buying it to check the AI box before they got fired, and I could see the box-checking more clearly than the product.

Then our agents started insisting on it. Repeatedly, until using anything else cost me more time than it saved. Part of that switch is the product being good. Part of it is that I have a finite number of hours and losing the same argument to an agent six times is a bad use of them.

My learning → Look at which tools your agents reach for without being told. That list is a buying signal your procurement process hasn’t caught up to, and it’s the list of companies with a distribution channel nobody can buy their way into.

The bull case for Clay being a $100 billion company

“The bull case is that agentic GTM has just started. We thought the TAM was the same as it was.

It turns out when agents can run these GTM motions, they will consume 10 to 100 times more usage than humans ever could. They can… https://t.co/FuWky90enI pic.twitter.com/tSwUjjMGsZ

— Harry Stebbings (@HarryStebbings) September 3, 2026

#4. 448 open tasks in a Replit build put me in Linear for the first time

I never needed project management. It’s me and the agents. Then the Replit task queue hit 448 open items and I couldn’t hold it in my head or in a doc.

Project management seemed to be a dying category. Humans don’t need Kanban cards and three-week handoffs anymore, and Asana’s performance says so. But the volume an agent-driven team generates needs a system of record, and Linear was built for that rather than retrofitted to it.

My learning → The tools you skipped as overhead when it was five humans come back when it’s one human and twenty agents. Reassess the categories you wrote off, because the bottleneck moved from producing the work to tracking it.

The Bull Case for Linear Being a $100BN Company:

"Linear is the clear winner.

They have built an agentic product first that allows us to build 100x more software, and that means 100x more features than ever before.

Humans cannot keep up with it, and humans still have to… pic.twitter.com/HG9cADSAah

— Harry Stebbings (@HarryStebbings) September 2, 2026

#5. Features that took a quarter now take a week, which makes your 2027 roadmap late

Not five minutes, and not the demo-video version, but a real week. Across the market that’s roughly 100 times more software getting built than 18 months ago.

The mistake I made on this show a year ago was sizing the coding TAM off the number of developers on the planet. The number of developers didn’t change. The amount each one ships did.

My learning → Pull up your 2027 roadmap this week. If it reads like a 2025 roadmap with more items on it, you’re planning at the old velocity. Your competitors’ roadmaps aren’t longer, they’re wider.

#6. Owner’s investors said it was too much software to build. Their CPO said there’s no choice. We’re all compound startups now

I sat in Owner’s board meeting last week, post $2.3B round. Their CPO went through the ship list and a room of experienced investors said this is too much. His answer was that they have no choice, this is the bar, and he doesn’t sweat that it’s ten times last year.

Their customers want the AI receptionist and the AI ordering and the rest of the stack. If Owner doesn’t build all of it, someone builds all of it and takes the account.

My learning → Decide now whether you’re building the suite or selling into someone else’s. Both work. Staying a point solution in a market where the adjacencies can reach you is a decision to be irrelevant in twelve months.

#7. Shipping 10x more product is how you end up with 40 buttons nobody uses

That’s what the investors in Owner’s board meeting were actually worried about, and they were right to worry. Volume is now easy to generate. Fitting all of that surface area into something a customer can use on the first try is not.

My learning → If you’re going compound, the product and design leadership hire moves ahead of the next five engineers on your list. Your agents will produce the features. They will not produce a coherent product.

#8. Cognition is third in coding at $46B and heading to $1.6B ARR

It’s reportedly raising at $46B, running $800M to $900M today. Sequoia used to say you couldn’t make money on the number three in a market. Then Postmates sold for a couple billion.

Cognition doesn’t need to catch Anthropic. $5B to $10B in ARR a few years out is a career.

My learning → Before you pivot away from a market because someone else is winning it, size what third place is worth. In a category growing this fast, third in the right market pays better than first in a small one.

#9. Iconiq’s data: companies growing over 100% grew headcount 133%

The 2025 thesis was that AI lets everyone run leaner. The data says something else at the top. Companies growing under 50% are adding no headcount and using AI for efficiency. The fastest growers are compounding software and humans at the same time.

That’s also why I think a lot of the more modestly funded companies in Europe are in trouble. Compound costs money, for tokens and for people, and you can’t out-compound someone with five times your balance sheet.

My learning → If your plan is flat headcount and win on efficiency, check what your fastest-growing competitor is doing with theirs. Against a competitor hiring into their growth, efficiency alone loses.

#10. Almost nobody on our team opens Salesforce anymore, and Benioff is fine with it