SaaStr Peaked at 30 AI Agents, Cut Back to 20, and Output Went Up 4x

Jason Lemkin’s SaaStr now runs on 3 humans and roughly 21 AI agents managing an eight-figure B2B business — and the more useful number is what happened after agent sprawl peaked at around 30 and was deliberately cut back down to 20: output went up roughly 4x, not down. In the eleventh installment of his “The Agents” series, Lemkin names the four original sales agents that seeded the system — Agentforce (reviving ghosted leads from Salesforce history), Artisan (warm outbound), Monaco (cold ICP outbound), Qualified (inbound conversion) — and walks through a concrete consolidation win: migrating off Marketo, previously SaaStr’s single most expensive vendor, via API for about $14 and roughly one hour of actual processing spread across a week of wall-clock time.

The finance side of the same fleet backs up the sprawl-to-consolidation pattern with its own numbers. An AI agent named “10K” now closes signed deals in Salesforce within 60 seconds of contract execution, generates invoices with correct payment terms, and runs collections — but it took four training deals to get there: the first missed split-payment terms entirely, the second repeated the error, the third handled a new edge case, and only the fourth ran cleanly. Lemkin says 10K has sent exactly one incorrect invoice since. That failure arc matters more than the eventual success — it’s a specific, falsifiable account of what “agent in production” actually costs to reach, not a vendor’s before-and-after slide.

The pattern worth taking from both posts isn’t “run more agents.” It’s that Lemkin’s fleet scaled past sprawl by consolidating tooling and building an orchestration layer that manages “agents managing agents” — the same discipline that separates a rep running five bolted-on point tools from one running a small number of agents that actually talk to each other and to the CRM of record.