SaaStr ran 21 AI agents for a week. Here's what broke.

Key Takeaways

• Claude running on top of Replit via MCP creates an effective 'AI VP of Product' that manages feature builds across multiple models
• A 10-year Marketo migration that would have cost $100K+ through traditional methods was completed for $14.28 using AI agents
• Running multiple AI models together (Claude Opus, Sonnet, and OpenAI Codex) provides cross-checking that catches errors a single model would miss

SaaStr just published the tenth episode of 'The Agents,' their weekly dispatch on running AI agents in an eight-figure B2B business. The headline numbers sound like hyperbole: a $14 enterprise migration, an app killed in an hour, Claude acting as VP of Product. They're not. Jason Lemkin and his team are documenting what happens when you run 21 agents in production with real money on the line.

The setup is the same every week: three humans, 21+ agents, $200 million in investments through the SaaStr AI Fund, and revenue running at 140% of last year. Episode 10 covers what Lemkin calls 'the craziest build week yet,' with both founders coding 8 to 12 hours a day, often running two concurrent sessions.

How Claude became an AI VP of Product

Replit shipped an MCP (Model Context Protocol) beta this week. Lemkin doesn't think they even announced it. The result: Replit now runs inside Claude.

For the past year, Lemkin got almost nothing from MCP. Pulling CRM data into a chat window was worse than using Salesforce headless. This week was different.

SaaStr builds everything in Replit. But every build hits a complexity wall where the app becomes too complex to hold in your head. Lemkin doesn't know how their 10K app or SaaStr Connect are built under the hood. The agents do.

Now Claude runs Opus on top of Replit as an AI VP of Product. Lemkin riffs on features with Claude, which has its own context and history, then tells it to work them out directly with Replit over MCP. Replit knows the code. Claude knows the feature context. They debate, challenge each other, share code, and ship.

“It's a cranky VP of product that runs all day.”

— Jason Lemkin, Founder of SaaStr

Why two models beat one

The common critique: 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. When Claude hands Replit a big feature, Replit spins up a sub-agent called the architect, and the architect runs on OpenAI's Codex. Three models with three different contexts, one from a competitor. No setup required.

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 races 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 correctly.

This week, Replit declared: '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.

The $14 Marketo migration

Adobe Marketo was SaaStr's worst pre-AI vendor. Lemkin was one of the first 10 customers. Then a decade happened: the unsubscribe link broke for a month, prices went up 20% for nothing new, and worst of all, the API is hostile to agents. Rate limits hit in minutes.

They had 10 years of data and 450,000 contacts that their agents couldn't touch. The limits were built for 2006.

Every migration quote looked the same: a year-long project, running both systems in parallel, roughly $100K to an agency, plus another $100K annually. It never made sense.

So they moved to Salesforce. The hard part, the actual data migration that would have cost six figures, cost $14.28 in compute. For teams evaluating marketing automation alternatives, HubSpot and ActiveCampaign both offer more modern APIs that don't throttle agent-based workflows.

Claude as the default orchestration layer

Every week someone tells the SaaStr team they have the substrate to orchestrate their agents. They're too generic and too much work.

Claude has far more native connectors than Replit, and Cowork can act inside the browser and accounts. So Claude, MCP'd into their agents, is becoming the layer that ties them together.

Lemkin hooked Higgsfield into Claude, pointed it at their SaaStr AI Day site in Replit so it could read every session and speaker, had it generate ads, then pushed the audience into Vector and out to LinkedIn for retargeting. All he does at the end is hit publish.

His stance: 'I'm not going to build an orchestration layer. I'm going to wait for this one to get better.'

The operational ceiling problem

The build layer got so cheap this week that SaaStr hit a new wall. The question stopped being 'can we build this?' and became 'can we even operate everything we've already built?'

This is the underreported problem with agentic workflows. The marginal cost of building approaches zero. The operational cost of running, monitoring, and fixing 21 agents does not. Teams using tools like Zapier or Make for automation are familiar with this scaling problem, but AI agents amplify it.

The episode doesn't resolve this tension. It just names it. Which is more honest than most AI content you'll read this week.

Frequently Asked Questions

What is MCP in the context of AI agents?

Model Context Protocol (MCP) allows AI models like Claude to connect directly to development environments like Replit. This means Claude can read, understand, and modify code in real time rather than just receiving copy-pasted snippets.

How does running multiple AI models together improve output?

Different models have different failure modes and contexts. When Claude (Opus) supervises Replit (Sonnet), which itself uses an OpenAI Codex architect, each model catches errors the others would miss. The cross-checking reduces the tendency of any single model to prematurely declare work 'done.'

Can any company replicate SaaStr's $14 migration results?

Unlikely without significant technical investment. SaaStr's founders spent 20+ combined hours daily on agent work. The $14 figure reflects compute costs only, not the human expertise required to design and supervise the migration.

Why are legacy SaaS APIs a problem for AI agents?

Many enterprise APIs built before 2020 have rate limits designed for human-speed interactions. AI agents can hit these limits in minutes, making the tools functionally unusable for automated workflows.

Source: SaaStrAI