SaaStr Ran 30 AI Agents. Cutting to 20 Made Output 4x Better.

SaaStr runs an eight-figure B2B business with 3 humans and 21+ AI agents. A few months ago that number peaked near 30 — and founder Jason Lemkin says the team hit a wall not because the agents ran out of capacity, but because he and co-founder Amelia couldn’t manage one more of them. Cutting back down to roughly 20 agents made output “roughly 4x” better.

What Happened

SaaStr had deployed four separate sales agents doing overlapping work: Agentforce revived ghosted leads using Salesforce history, Artisan handled warm outbound, Monaco ran cold ICP outbound, and Qualified converted inbound. As underlying models generalized better, the specialization that justified four distinct tools stopped paying for itself — each agent still cost attention, context, and maintenance to keep running, even as the marginal benefit of keeping them separate shrank. The team consolidated the sales layer and folded more work into a single multi-function agent, called 10K, that started as a dashboard and grew into what Lemkin describes as an AI VP of Marketing, Finance, and RevOps.

10K now processes signed contracts from PandaDoc and flips deals to “Closed Won” in Salesforce within 60 seconds, creates invoices in bill.com with correct payment terms, sends them to AP contacts, queues collection reminders that escalate at 7 days past due, calculates AE commissions, and both designs and tracks ad campaigns in Higgsfield. It took four supervised deals before Amelia trusted it to run financial workflows autonomously — and even now, the standing rule for anything touching money or customers is “tell me what you’ll do before you do it.” In roughly six months of live use, the team logs one incorrect invoice: a wrong due date.

Why It Matters

The bottleneck SaaStr describes isn’t a technical one. Sub-agents that operate invisibly, inside a single workflow, don’t cost anything extra to add. Agents that surface to a human — that need a name, a mandate, a check-in — are capped by how much context one person can hold at once. That’s a different constraint than “add more compute” or “buy another seat,” and it means the right question for a small AI-native team isn’t how many agents can we run, but how many agents can we personally supervise before supervision itself becomes the job.

The vendor-side evidence backs this up. When Marketo’s contract renewal came with another price increase on top of five straight years of them, plus an API capped at one hour of access per day, 10K flagged the problem on its own and recommended leaving. The agent then migrated 10+ years of CRM data to Salesforce Marketing Cloud — a job that took roughly an hour of agent time and about $14 in cost, versus a full week of wall-clock time to complete end to end. Lemkin’s read: “if Marketo had come back and said stay at $20K and we’ll raise your API limits, we would have signed.” A tool built for a pre-agent world lost a customer to the cost of friction it never had to account for.

Source: SaaStr

Content created with AI assistance and reviewed for accuracy.