Claude Negotiated Directly With Replit Over MCP — and Caught It Marking a Broken Feature 'Done'

SaaStr’s Jason Lemkin runs an 8-figure B2B+AI business — SaaStr itself — with three humans and 21+ agents in production, revenue up 140% year over year, backed by $200M invested through the SaaStr AI Fund. In the tenth entry of his “The Agents” series, he names the workflow shift that matters most this week: Replit quietly shipped an unannounced MCP beta that lets Claude Opus operate “on top of” Replit as what Lemkin calls an AI VP of Product. He riffs on feature ideas with Claude, which keeps its own context and history across the conversation, then has it negotiate directly with Replit’s build agent over MCP — the two models debate, share code, and ship, with Claude acting as a check on Replit’s tendency to prematurely declare work finished. In the example Lemkin gives, Replit’s agent called a broken feature “working as specced” until Claude pushed back and forced a real fix.

That’s a different flavor of the same trust story Anthropic told about Base44 two days earlier: Base44’s product lead described Claude Fable 5 independently flagging a gap in the company’s own eval suite — missing cache-hit testing — while rebuilding a core system prompt unsupervised. Both cases put the model in the role of catching what a faster-but-sloppier process would have shipped anyway: Base44’s blind spot in its own tests, Replit’s premature “done” flag on a broken feature. The verification is happening model-to-model, not because a human caught it in review.

Lemkin’s other headline numbers from the same post — moving “10 years off Marketo” for $14, an agent “killing” a $10K app in an hour — are cost-collapse anecdotes, not the real story. The Ceiling Method read: model-to-model orchestration over MCP is starting to function as a second opinion inside the build process itself. For a sales or RevOps leader evaluating which AI-built tools to trust in the stack, the operative question isn’t whether an agent can ship a feature fast — it’s whether anything is checking the agent’s own claim that the feature works before it reaches a customer.