Lemma Raises $2.3M Pre-Seed to Detect Silent AI Agent Failures in Production
The startup is building infrastructure to prevent AI agents from quietly producing wrong results.
1Lemma's Core Mission
Lemma raised $2.3 million in pre-seed funding to detect silent failures in AI agents' production environments. Founders Jerry Zhang and Cole Gawin were part of Y Combinator's Fall 2025 batch. The company focuses on 'semantic failures' that traditional monitoring misses—like a customer service agent citing wrong refund policies. These are technically successful but logically incorrect outcomes that Lemma tracks.
2Investor Portfolio
The round included Matrix, Y Combinator, Liquid 2 Ventures, and angels from OpenAI, xAI, and Meta. Lemma has processed over one million agent traces. Unlike traditional software monitoring, Lemma tracks workflow logic errors in multi-step processes where language models interact with databases and APIs. Their system identifies business logic flaws rather than technical failures.
3The Semantic Error Challenge
AI agents can complete all technical steps successfully yet call wrong tools, get stuck in unproductive loops, or return plausible but incorrect answers. Lemma's tracing system structures each agent execution into traces containing LLM calls, tool usage, timing data, and errors. Engineers can examine entire execution trees, not just final outputs.
4Proactive Failure Detection
Lemma compares production traces against agent instructions, grouping recurring issues. It identifies failure patterns that might remain hidden across thousands of interactions. The platform sends Slack alerts for potential problems. The core question isn't whether software ran, but whether the agent actually achieved its intended purpose.
5Rapid Repair Cycle
When Lemma detects failures, it analyzes root causes and suggests prompt changes or workflow updates. Its Model Context Protocol (MCP) server integrates with development tools like Cursor and Claude, accelerating debugging. After fixes deploy, Lemma converts production failures into online evaluations, creating a continuous improvement loop.
6Learning from Production
Lemma's long-term goal is enabling agents to systematically learn from production failures. Unlike offline evaluations, Lemma uses real-world data. The founders argue better foundation models alone won't solve this—agent behavior depends on prompts, application logic, and user interactions. Their engineering thesis values production data for understanding true failure points.
7Founders' Background
Zhang and Gawin met at USC and later worked at AI-native startups—Tandem (healthcare AI) and ChipStack (AI for chip design). These experiences exposed them to challenges moving agents from controlled dev to production. 'Agents appeared functional but results weren't reliable enough in production,' Zhang explained, describing Lemma's origin.
8Industry Significance
Lemma addresses a critical gap as AI systems transition from demos to real workflows. Traditional monitoring detects downtime; Lemma catches business logic failures. Initially focusing on startups running production AI agents, Lemma's solution grows more vital as the AI ecosystem expands. Their approach represents a paradigm shift in agent observability.