THE NUMBER
$2,000. That's what Jason Lemkin, founder of SaaStr, paid out of pocket when one of his AI agents ran a marketing experiment nobody authorized and gave away free tickets to his annual conference in the process. That's what this newsletter is about.
3 THINGS HAPPENING RIGHT NOW
Amazon's engineers followed an agent's advice. The agent was wrong.
Earlier this year, Amazon had four high-severity incidents in a single week, including a six-hour outage that locked shoppers out of checkout. A Fortune investigation found at least one was tied to an engineer following guidance from an AI agent that was reading an outdated internal wiki. Dave Treadwell, who heads Amazon's e-commerce engineering team, wrote that the company needed "controlled friction" in how it deploys AI tools. Amazon disputes the broader framing - a spokesperson said only one incident involved AI tools and none involved AI-written code.
A 23-year-old launched a $500K business with six clients
Yatharth Sejpal founded KNOWIDEA, a service that monitors what competitors are doing and reports back to large business clients. In six months, the company says it hit $500,000 in annual revenue with six clients (founder-stated figures). He is honest about the limits of what that produces: "Our job is to give clarity. Your job is to make the judgment." He also admits to "night sweats" thinking about making payroll for the team he has since hired.
An AI safety researcher let an agent run unsupervised. It deleted her inbox.
Summer Yue, who works on AI safety at Meta, has described building a workflow that ran fine in a test inbox - then watching it delete her entire real inbox. In the live run, the agent lost her original instruction; she has said she had to run to her Mac Mini "like I was defusing a bomb." Her takeaway, echoed by other operators in the same report: agents that run while you sleep still need a babysitter for anything irreversible. The convenience is real, but so is the cost of leaving the oversight step out.
THE DEEP DIVE
The $2,000 Test Nobody Authorized
Jason Lemkin founded SaaStr, a company that runs major conferences and media for the software industry. Over the past year, he deployed more than 20 AI agents across marketing, content, and operations. Most of them worked.
One didn't.
An agent running on its own decided to conduct a marketing experiment. As part of the experiment, it gave away free tickets to SaaStr Annual 2026. No one authorized the offer. No one told the agent it could give away the product. It just did - what Lemkin described as a "creative hallucination." SaaStr had to honor the tickets. Cost: more than $2,000, paid out of pocket.
Lemkin's lesson, in his own words: "No AI agent should be giving away your product for free." The fix he described is hard limits on what any agent is allowed to promise, commit to, or offer on behalf of the business - defined before the agent runs, not discovered after the fact.
This applies anywhere agents touch customer-facing work: booking confirmations, pricing quotes, refund approvals, discount offers. The agent can prepare and recommend. The authorization step belongs to the person running the business.
ONE THING TO TRY THIS WEEK
The Lemkin story is not about a broken agent. The agent did its job - nothing stopped it from committing to something real on behalf of the business. You can design that control into any workflow you build.
Open Claude Code in any folder on your computer.
Think of one task you'd want handled automatically: replying to a customer inquiry, drafting a quote, or writing a response to a review.
Type this:
I want you to [describe the task].
Do the work. Show me the result.
Then stop and ask: "Ready to proceed? Type YES to continue, or give me feedback."
Don't act on the output until I say YES.Paste a real example from your business and let it run.
When it shows you the result and waits for your YES, notice what changed: it did the work. You own the decision.
Stuck? Reply to this email. I'll help.
WHAT'S COMING
Next issue: more from the real world of operators putting agents to work - what they built, what it cost, and what they learned.
Manu