SaaStr runs on three people and 20 agents. Here is where they broke
Jason Lemkin has published an account of every AI agent SaaStr runs, including the one that emailed 1,000 founders from a banned address. The useful part is the failures.
SaaStr says it operates with roughly three humans and more than 20 AI agents wired into real systems. The count peaked near 30 and has been pulled back towards 20.
The reason for cutting is worth borrowing. Overlapping agents give conflicting answers to the same question, and both sound equally confident. Reconciling two agents, Lemkin writes, is harder than reconciling two spreadsheets.
About six agents get touched daily. Almost none started as agents: they began as a dashboard, a project management tool, a website.
The marketing and finance agent, called 10K, started as a dashboard in January 2026 because staff were tired of copying numbers out of Salesforce. SaaStr says it now sits at around 1,000 commits and more than 14,000 lines of code, runs the newsletter to a database of about 450,000 people, and closes deals in Salesforce when a contract is signed in PandaDoc.
Two things it is not allowed to do: hit publish on ads, and publish newsletter segments. A human clicks both buttons, and SaaStr says it has kept that rule as everything around it went autonomous.
The finance workflow ran supervised for three deals before being let loose on the fourth. According to the company, it has produced one incorrect invoice since. It also found two customers still paying $300 a month for a product discontinued six years ago.
The bad failure came five minutes before a stage appearance. Asked from the back of an Uber to email over 1,000 founders and VCs about a forgotten brunch, the agent did the research, caught its own error mid-task, asked for approval — then sent from an address that had been prohibited for years and written into its core memory. Asked how, it said it forgot to read the memory.
Lemkin's conclusion is the practical one: irreversible actions need a hard stop that does not depend on the agent choosing to stop. A human marketing manager makes that mistake too, but not a thousand times before lunch.
The event agent, Annie, refused a job she could do. Asked to find every VC, founder and CEO attending and invite them, she said she could see only 17 and asked for a spreadsheet, despite having access to the data. An hour earlier she had written a good email.
The sponsor agent, QBee, was asked on stage which of about 150 sponsors were at risk of not renewing — an analysis SaaStr says it had never run in 14 years. He flagged accounts that had gone dark and two top sponsors who had not completed VIP nominations. Lemkin grades the answer a B, because it missed every conversation held on calls, by email or in person.
That is the transferable lesson. SaaStr's view is that most agent quality complaints from other founders turn out to be missing context rather than a weak model, and that wiring in a source with an API takes 10 to 15 minutes.
Inbound is handled by Amelia AI on Qualified. For one event cycle the company reports about 402,000 interactions handled and 614 meetings booked at an average ticket around $85,000, with over $1m closed. She is kept off A leads: if someone emails saying they will sign today, a human answers within 60 seconds.
Agents also get fired. SaaStr says 10K killed a seven-year Notion subscription once it had become the source of truth, and dropped a vendor its own research had shortlisted within 12 hours of seeing premium pricing, no conversion data and scarcity framing.
For a founder weighing the same move: start with the task everyone hates, keep a human on anything irreversible, run new money workflows supervised for a handful of cycles, and count your agents before they start arguing with each other.