Key Takeaways

  • Lemma raised $2.3 million in pre-seed funding to catch AI agent failures that never trigger an error, the kind that quietly cost companies their customers.
  • Founders Jerry Zhang and Cole Gawin left the University of Southern California to build Lemma full-time, after a short planned break from school turned permanent.
  • Lemma’s software now watches more than a million agent traces a day, flagging problems like fabricated ticket numbers and routing a fix straight to the developer’s coding tool.

Jerry Zhang and Cole Gawin left the University of Southern California to build software that watches AI agents nobody else is watching. Their company, Lemma, just raised $2.3 million in pre-seed funding. Backers include Matrix, Liquid 2 Ventures, and operators at OpenAI, xAI, Meta, and DoorDash. Forbes also recently named Lemma one of the top startups to watch from Y Combinator’s Fall 2025 batch. 

The problem is easy to miss. Most teams monitor AI agents the way they monitor ordinary software: they watch dashboards for crashes, errors, and downtime. But an agent can fail while everything still looks green. It can loop, invent information, or misunderstand a request, then return a response marked successful. The customer quietly gives up, while the monitoring system never sounds an alarm.

The idea behind it is simple to describe and hard to solve. An AI agent can fail without any of the usual warning signs. Nothing crashes, no error appears, and uptime never dips. The customer just quietly gives up on the other end.

Leaving USC to Build What They Needed

Zhang and Gawin met as students in SEP, USC’s startup program. They arrived there from very different directions. Zhang grew up in Canada, winning almost every competition he entered. He earned a perfect GPA, competed in national math contests, and ranked fourth in the country in the 200-meter breaststroke. Gawin taught himself to code in suburban Chicago at age nine. He built a website for his soccer team on Weebly, then wondered who had built Weebly itself. That question led him to PHP while he was still in elementary school. He worked engineering internships in high school, then turned down a job offer for college. Once there, he chose cognitive science over computer science. Anyone can learn to code from YouTube, he reasoned, but nobody can run a brain lab at home.

Their first product together, Clinicode, handled medical insurance paperwork. The idea traced back to Gawin’s mother, a physical therapist who filed claims each night at her kitchen table. By age 18, the pair were selling to hospital staff around Los Angeles. They won a $15,000 grant from the university and became the first team there to win both Best Product and Best Demo. Clinicode folded soon after. Running a startup alongside a full course load asked more than either had to give. Both had planned only a short break from school to build what came next. That break never ended, and neither has been a full-time student since.

Two Internships, One Shared Problem

They found their next idea a summer later, from two different offices. Gawin was interning at a healthcare AI startup, where he spent his time rewriting prompts by hand. He would change a word, run the test, read the result, and start over. After a few weeks, he saw that a program could handle the whole cycle. He built one and wrote himself out of the role. When he told Zhang about it, Zhang already understood the problem well. Engineers at his own internship were stuck in the identical loop. Two people hitting the same wall at unrelated companies suggested the problem reached far beyond either desk.

An investor who had once passed on Clinicode sent them $30,000 to try again. Both took a semester off to build. Zhang talked Gawin into applying to Y Combinator, a bet Gawin resisted at first. Their interview lasted ten minutes and turned on a single question. A partner asked how their approach differed from DSPy, a research tool Gawin had read that same morning. He answered in detail, then opened his laptop for an unplanned demo. The acceptance call came the following day.

Turning Silent Failures Into a Product

During the program, the founders worked alongside dozens of companies already fixing AI systems in daily use. The same complaint surfaced again and again. Monitoring tools built for ordinary software could not keep pace with AI agents. What looked like a detour from building their own product turned out to be their sharpest advantage. The two watched firsthand where agents actually break. Gartner now predicts that more than 40% of business projects using AI agents will be canceled by the end of 2027. The firm points to rising costs, unclear business value, and weak risk controls.

“Cole and I started Lemma because we experienced the pain of building AI agents firsthand,” said Zhang. “We kept running into the same problem: agents would appear to work, but the results weren’t reliable enough in production. We wanted to build the tools we wished we had: something that helps teams catch issues earlier, learn from production data, and continuously improve agent performance in the real world.”

Lemma’s software watches an agent’s live traffic and traces each failure back to its root cause. It then messages the team in Slack with the damage attached, such as a note that fabricated IDs are affecting two-thirds of runs. A suggested fix goes straight to the coding tool the developer already has open. To make the problem tangible, the company paired its funding news with a launch video. It recreates a viral restaurant scene, where the order looks handled right up until it arrives wrong. Lemma’s software now watches more than a million agent traces a day. Its customers run from seed stage through Series B.

To see how it works or book a demo, visit the official site.