About Lemma
Lemma is production monitoring infrastructure for AI agents that detects silent failures, where an agent appears to succeed but does the task wrong. It is part of YC's Fall 2025 batch. In August 2026 it raised a $2.3M pre-seed round with Matrix and Y Combinator.
What is Lemma?
What is Lemma?
Lemma focuses on a specific gap in AI observability: failures that do not throw errors. An agent can return a response, finish every tool call, and still get the task wrong. Traditional observability tools that watch latency and exceptions miss these semantic failures.
According to the company, Lemma scans every trace to surface issues before users complain, identifies root causes, and helps teams fix them without manual digging. The SaaS News reports that the platform captures structured traces of agent executions, including LLM calls, tool usage, inputs, outputs, retrieval steps, timing data, and errors, then analyzes them to find recurring failures, prioritize issues, and alert teams.
Lemma was founded by Jerry Zhang and Cole Gawin and is based in San Francisco. It was part of Y Combinator's Fall 2025 batch. In August 2026 it announced a $2.3M pre-seed round with participation from Matrix, Y Combinator, Liquid 2 Ventures, Vermilion Cliffs Ventures, Irregular Expressions, Cervin Ventures, Comma Capital, Position Ventures, Eight Capital, and operators from OpenAI, xAI, Meta, and DoorDash. The company said its initial focus is startups already running AI agents in production.
Pros
- Targets semantic failures that standard monitoring misses
- Captures structured traces across LLM calls, tools, and retrieval steps
- Prioritizes recurring issues and points to root causes
- Built specifically for agents running in production
Cons
- Most useful once an agent is already in production with real traffic
- Pricing is not listed publicly
- Overlaps with broader LLM observability platforms teams may already use
How to Use Lemma
- 1Sign up
Request access on the Lemma website.
- 2Instrument your agent
Send agent traces, including LLM and tool calls, to Lemma.
- 3Review detected issues
See recurring semantic failures ranked by priority with likely root causes.
- 4Fix and monitor
Ship fixes and keep monitoring so regressions are caught before users report them.