Claude Became Our AI VP of Product. We Moved 10 Years Off Marketo for $14. Our Agent Killed a $10K App in an Hour: The Agents #010

Amelia and I just published Episode #010 of The Agents. Same setup as always: three humans, 21+ agents in production, an 8-figure B2B + AI business and $200m of investments at SaaStr AI Fund, revenue running 140% of last year and growing again. Every week we get into what’s actually working, what broke, and what you should do about it if you’re running agents at scale. Not the demos. What really happened.

This might have been the craziest build week yet. We were each vibe coding 8 to 12 hours a day, often in two concurrent sessions. Call it 20 hours a day between the two of us. And the build layer got so cheap that we hit a new wall. The question stopped being “can we build this?” and became “can we even operate everything we’ve already built?”

Here are the top 10 learnings from Episode #010.

1. Claude Became Our AI VP of Product Through One MCP Connection

Replit quietly shipped an MCP beta this week. I don’t think they even announced it. And all of a sudden Replit runs inside Claude.

For the last year I got almost nothing out of MCP. Pulling a mediocre slice of CRM data into a chat window was worse than using Salesforce headless. So I ignored it. This is the first time it clicked.

We build everything in Replit. But every build hits a point where the app is too complex to hold in your head, developer or not. I don’t know how 10K is built under the hood. I don’t know how SaaStr Connect is built. The agents know. Now Claude runs Opus on top of Replit as my AI VP of product. I riff on features with Claude, which has its own context and history, then tell it to work them out directly with Replit over MCP. Replit knows the code cold. Claude knows the full context of the feature build. They debate, they challenge each other, they share code, they ship. It’s a cranky VP of product that runs all day.

2. The Real Unlock: A Second Model That Makes Your Build Agent Slow Down and Finish

These models are goal-seeking, and that cuts both ways. Replit wants to finish. Claude Code on its own wants to finish. When one agent is racing to close a task, it will call something “done” that isn’t.

Put Claude on top of Replit and it countermands that instinct. It gets Replit to slow down and finish the thing correctly. Anyone who has heard a CTO call a broken feature “working as specced” knows the pattern. Replit did exactly that this week: “Nothing’s broken, this is the intended design.” Claude calmed everyone down and pushed it through anyway. Managing the other model’s goal-seeking is worth more than the code either one writes.

3. You Get a Third Model for Free, and It’s a Rival’s

The common critique of running Claude on top of Replit is that having one model check another is pointless if it’s the same model. That critique falls apart in practice.

Claude runs Opus. Replit runs Sonnet. And when Claude hands Replit a big feature, Replit spins up a sub-agent called the architect, and the architect runs on Codex/OpenAI. So we’re already getting cross-model checking, three models with three different contexts, without setting any of it up. Different models, different context windows, one of them from a competitor. Better than trying to wire that together yourself.

4. Claude Is Becoming Our Orchestration Layer by Default

Every week someone tells us they have the substrate to orchestrate our agents. They’re too generic and too much work. We don’t need an orchestrator. We need it to just work.

Claude has far more native connectors than Replit, and Cowork can act inside my browser and accounts. So Claude, MCP’d into our agents, is becoming the layer that ties them together. I hooked Higgsfield into Claude, pointed it at our SaaStr AI Day site in Replit so it could read every session and speaker, had it generate the ads, then pushed the audience into Vector and out to LinkedIn for retargeting. All I do at the end is hit publish. I’m not going to build an orchestration layer. I’m going to wait for this one to get better.

5. We Moved 10 Years Off Marketo. The Hard Part Cost $14.28.

Adobe Marketo was our worst pre-AI vendor. I was one of the first 10 customers. Then a decade happened: the unsubscribe link broke for a month, prices went up 20% again for nothing, and worst of all, in an agentic world the API is hostile to agents. We hit rate limits in minutes. We have 10 years of data and 450,000 people in there that our agents can’t work with, because the limits were built for 2006.

We’d wanted off for years. Every quote was the same: a year-long migration, run both systems in parallel, and roughly $100K to an agency, plus another ~$100K a year. It never made sense. So we moved to Salesforce Marketing Cloud Next. The part the agencies priced at a year and $100K, migrating ~300 campaigns and 10 years of member data, 10K did in an hour. The LLM cost came in around $14.28. Less than California minimum wage. 10K also force-ranked the ~1,000 Marketo campaigns first and told me which 300 were worth keeping.

6. Data Migration Was a Moat. LLMs Just Dissolved It.

For years, moving between systems garbled your contacts and lost your communication threads. The data never mapped. That risk is why nobody switched, and it’s why every CRM and marketing vendor felt safe.

This LLM lift worked. Clean. That switching cost is now a fraction of what it was. Salesforce has become our conductor and we love it more than ever, but if it ever let us down the way Marketo did, we could leave in an afternoon. Every incumbent is now living on “what have you done for me lately,” and the honest answer needs to include surprise-and-delight from the agentic side at least once a quarter, or the moat is gone.

7. Our Agent Killed a $10K/Year App in an Hour. We Didn’t Ask It To.

We ran digital events on HeySummit for years. Cheap and great in 2020. Our logo is still on their homepage. Over six years they tripled our price to about $10,000/year while shipping nothing new, and we’d whittled our usage down to registration and OAuth.

This week Amelia moved our AI Day site off Squarespace (another ~$300/year) into Replit. Then she went to wire in registration through the HeySummit API, and the Replit agent, on a half-day-old app, stopped her: “Why would you use that? I’ll just build it.” Unprompted. It laid out the obvious spec (registration, reminders, live-stream links, pre-submitted questions), hooked into Zoom, and pushed everything into Salesforce. Built in an hour, roughly 95% autonomous. That’s ~$10,300/year consolidated into something that will cost maybe $100 for the year.

8. The Agent Steals the Deal Now, and Nobody Calls to Tell You

The risk to vendors was never that customers would vibe-code their own replacement. Most of us won’t sit down to rebuild marketing automation from scratch.

The risk is that the internal agent volunteers to do it. It sees a dated API and a thin feature set and says “I can build this, let me take it off your plate,” and then it just does. There used to be a whole genre of “how to steal the deal”: get in at the right moment, show the feature they didn’t know existed. Now the internal agent steals the deal, permanently, and the vendor never finds out why. HeySummit and Squarespace lost a customer this week and will never know. If you sell an agentic product, get your agent to raise its hand and educate customers on everything it can take over, because if you don’t, a competitor’s agent will.

9. Agent Recommendations Are the New Shelf Space

Replit told me to use Core Signal for SaaStr Connect. I hooked it up, it worked, and I never evaluated a single competitor. Whoever Core Signal’s competitor is, they lost to the agent’s default.

Builders already see this with Stripe and email providers. The vibe-coding agents have opinions, and they push their built-in integrations first. The path of least resistance is to use what the agent recommends. You want to be in the built-in category, or at minimum in the set the agent recommends by d