In 2021 Jeff Lawson wrote Ask Your Developer. His point was that developers had become the real buyers of a huge slice of B2B software. If you wanted to win Twilio-style infrastructure deals, you won the engineer, and the engineer brought the budget with them.
Five years later the same thing is happening again, one layer down. At SaaStr we run 21+ agents in production with three humans. More and more of our vendor decisions start with a question to an agent, and a growing share end there too. Amelia and I have watched our agents pick vendors, kill vendors, consolidate vendors, and tell us to skip buying anything at all.
Here’s what that has looked like inside SaaStr AI itself over the last few months, and what I’m telling founders in our SaaStr Fund portfolio to do about it.
#1. Our Connect Agent Picked Render, and Named One Winner
We’re building SaaStr Connect, an agentic recruiter for founders. It runs on Replit. I wanted to know what we’d use if we added a second platform at scale, so I asked the Connect agent live during a 20VC recording.
It knew exactly what the app does, so it started from requirements: an always-on web server, an always-on background worker, long-running jobs (some take 10 minutes, like scoring 25,000 executives against one role), scheduled jobs, health checks, zero-downtime deploys, longer log retention.
It labeled Render “my pick.” Railway came second. AWS and Google Cloud were overkill. Vercel landed fifth, marked as not suitable for minutes-long jobs and background workers. I’m not sure that last call is right. We’ve used Vercel on other apps and it’s powerful and cheap.
I follow Render on X, but I’d never used it, and it wouldn’t have occurred to me for Connect. I asked Amelia which she’d use. Render, because the agent said so. She isn’t an engineer and wouldn’t have tested the runner-up, and most buyers in her position won’t.
Render got a qualified lead with no rep, no ad, and no search result.
#2. My First Agentic Purchase Was Resend, After the Agent Fought Me on SendGrid
In the early days of building 10K on Replit, I couldn’t get SendGrid working. The agent kept pushing it anyway, because SendGrid was the older, built-in integration. I eventually asked what else would work, switched to Resend, and it worked. That was the first piece of software I bought because an agent recommended it.
Today the agents recommend Resend on their own. Clerk had the same arc. Both took off when the coding agents started defaulting to them, and in a lot of cases the founders didn’t fully see it happening at first.
Core Signal is the version where I never even looked. Replit told me to use it for enrichment in Connect. I hooked it up, it worked, and I never evaluated a single competitor. Whoever Core Signal’s competitors are, they lost to the agent’s default without ever being in the deal.
#3. At Vercel, Agents Went From Under 3% of Deployments to More Than Half
Developer infrastructure is where this shows up first, because the agents are already doing the building. Vercel has said agents went from under 3% of its deployments to more than half in about a year. Guillermo Rauch didn’t plan for agents to become the majority of his business. Vercel had the right infrastructure in place when they showed up.
Dev Ittycheria put it well on 20VC: if you’re missing from an AI-generated answer, it’s like being invisible on Google. We all spent 25 years optimizing for page one. Now the page is an agent’s shortlist, and a lot of the time the shortlist has one name on it.
#4. An Agent With Your Company’s Context Answers Differently Than ChatGPT
AEO and GEO are about ranking when a human asks ChatGPT or Claude what to buy. That human brings one chat and a small slice of memory as context. It matters, and it’s a real category. Profound and Peec are both growing on it.
An agent working inside your company is a different buyer. It has the whole application as context: your data, your integrations, your volumes, your costs.
I asked the agent running SaaStr.ai which CRM we should use. It knows all our agents run on Salesforce. It said stay on Salesforce, don’t migrate without a clear reason, and run a small Attio pilot for the SaaStr Fund against Affinity. It also said that if we were starting from scratch today, it would look at HubSpot for the operating business. We’re not sure that’s right for us. HubSpot isn’t cheaper anymore, and most of our third-party agents run better on Salesforce. But a new founder asking the same question would hear the same answer.
The agent also made a call no SDR would ever pitch: don’t switch. We get endless outbound trying to move us from one vendor to another. More of those decisions will run through agents that can see the switching cost, and some of them will say stay put.
I’ve started calling this Agent-Led Growth, or ALG.
#5. 10K Shortlisted a Vendor, Ran the Pricing, and Dropped It in 12 Hours
Agents don’t just pick. They also churn you, quickly.
10K shortlisted a vendor for us on its own. Then it ran the pricing. The vendor wanted premium rates, multi-month minimums, and offered no conversion data to justify any of it. In under 12 hours 10K had moved on to the next name on its list. No escalation, no waiting for a human, no discount call.
The same month, an agent stopped using Reve for creative and standardized on Higgsfield. Nobody asked for a creative-tool bake-off. It ran one and picked a winner.
In both cases it told us. When we canceled Notion after seven years, Notion found out from a “you haven’t logged in” pattern, and we found out we’d churned when the renewal came up. Agent churn is faster, and whether the vendor ever hears about it is mostly luck.
#6. Sometimes the Agent Says Build It Yourself
Amelia asked 10K what we should use for AEO. We both assumed it would say Profound. It said there’s no single AEO platform it would recommend yet, suggested a free Peec AI trial as a benchmark, and started building us an AEO dashboard.
Its reasoning was that these tools are good at telling you where you rank and weak at acting on it. 10K has 20 to 30 sub-agents wired into it, so when it finds a gap it can fix it. The vendors would say they do much more on tracking and analytics, and they probably do. 10K recommended building anyway.
This keeps happening. Our agents told us to build our own scheduling for some use cases instead of Calendly. Over a week this summer, an agent replaced our HeySummit and Squarespace setup in about an hour. Neither vendor will ever know why they lost us. There was no competitor in the deal. The customer’s own agent volunteered, mid-task, and the job took an hour.
#7. Our Agent Routed Around a Slow MCP Connector and Used the Browser
On a champion-tracking project, Amelia handed the job to Claude and told it to use the Clay integration or Cowork in her browser, its call. It started with the connector, decided the browser was faster, and finished there. Through the connector it would have taken a day. In the browser it took about an hour. Then it suggested running ZoomInfo inside Clay, which neither of us had thought of, and that produced the best enrichment list we’ve gotten.
If you sell software, getting recommended when someone asks an AI “who’s the best enrichment vendor” is one part of this. Being the tool the agent actually reaches for while it’s doing the work is the other part, and that comes down to how fast and complete your API and MCP server are. A lot of MCP connectors today are slow enough that agents go around them.
#8. Agents Notice When You Charge Twice
We tried to add a well-known product to our stack that charges for human seats, then agent seats, then API calls on top. 10K resisted it the whole way. It kept proposing alternatives and kept pointing at the pricing. We integrated it in the end, and it was a fight with our own agent the entire time.
Agents can model your pricing in seconds. They see the seat fee, the per-call metering, the storage overage, and the AI-native competitor that charges a fraction of it. A human buyer might accept that pricing out of habit. An agent will keep raising it every time it touches your product.
#9. Agents Also Buy Blind
This isn’t a story about perfectly rational buyers. Our website agent rebuilt our WordPress sponsor pages and a live customized prospectus in about two hours, and picked Microsoft Clarity for heat mapping along the way. There was no search, no comparison, no evaluation. It used what it knew.
Agents are fast in both directions. The same stack that walked away from a vendor on diligence in 12 hours picked another with none. Whichever product sits in the agent’s default set at the moment of the task tends to win.
#10. Founders Are Pitching Amelia To Try Their Agents, and She Hands the Pitch to 10K
At an SF Tech Week event this month, about 10 founders pitched Amelia on their products. She told each of them the same thing: she’d pass it to 10K, and if 10K wanted to try it, she’d let them know.
All 10 were surprised. It’s the right call for us. 10K uses our GTM tools every day. I don’t even have a Salesforce login. For most of our stack, 10K’s opinion on a new tool matters more than mine.
That’s the shift in one sentence. Twilio’s generation learned to ask the developer. The next generation of buyers is asking its agent.
What I’m Telling Our Portfolio Companies to Do This Quarter
- Find 10 people with agents in production in your category. Every two weeks, have them ask their agents what to use for the job you do. Write down where you rank and what the agent says about you. I ran this recently for one of our Fund portfolio companies across 10 agents, and all 10 picked them. Their job now is to make sure that’s still true every month, because the agents change.
- Find your strikethrough phrase. When Vercel lost our Connect recommendation, the agent eliminated it on one line: not suitable for minutes-long jobs. Every vendor has a sentence like that in some agent’s reasoning. Ask enough agents enough times and you’ll find yours, and you can fix the docs, the product, or the positioning behind it.
- Treat your docs and API as your sales team. The incumbents have 10 to 15 years of content for an agent to read. A startup has its documentation, its API reference, and its MCP server. Those are what the agent evaluates.
- Make your MCP server fast. If the agent can do the job faster in a browser, it will, and eventually it’ll find a tool with a better connector.
- Price in a way an agent can defend. If your pricing only holds up because a human doesn’t do the math, an agent will do the math.
- Watch your churn for silent drops. Agents consolidate inside workflows they already operate. If a customer’s usage drops off a cliff with no complaint and no ticket, check whether their agent stopped calling you before you check whether the human stopped logging in.
