Lemma Raises $2.3M Pre-Seed to Build Monitoring Infrastructure for AI Agent Failures

AI agent reliability startup Lemma has raised $2.3 million in pre-seed funding from Matrix, Y Combinator, and others to build monitoring infrastructure. Founded by Jerry Zhang and Cole Gawin, Lemma addresses "silent" or semantic failures where agents technically execute workflows but produce incorrect results. The platform has processed over one million traces, helping developers debug agents and mitigate risks like "excessive agency" by tracking tool calls, security signals, and task outcomes.

Lemma $2.3M pre-seed funding

• ▪The AI agent reliability startup Lemma raised $2.3 million in pre-seed funding to build monitoring infrastructure for detecting silent AI agent failures.
• ▪Lemma was founded by Jerry Zhang and Cole Gawin, who were part of Y Combinator's Fall 2025 batch and previously worked at Tandem and ChipStack.
• ▪Lemma's pre-seed funding round included participation from Matrix, Y Combinator, Liquid 2 Ventures, Vermilion Cliffs Ventures, Irregular Expressions, Cervin Ventures, Comma Capital, Position Ventures, and Eight Capital.

Silent AI agent failure problem

• ▪Traditional software monitoring is designed to detect explicit technical failures like application crashes, error codes, latency spikes, or infrastructure downtime.
• ▪Lemma describes silent agent failures as semantic failures, which include customer service agents citing incorrect policies or auditing agents generating outdated reports.
• ▪AI agents introduce silent failures where an application remains technically operational but the agent misunderstands user intent, calls the wrong tool, or produces incorrect results.

Semantic failure detection infrastructure

• ▪Lemma integrates with development environments through a Model Context Protocol server, allowing developers to query production traces from tools like Cursor, Claude Desktop, and Claude Code.
• ▪Lemma's tracing system converts agent executions into structured traces containing large language model calls, tool invocations, inputs, outputs, timing data, and retrieval steps.
• ▪Lemma's platform has processed more than one million agent traces to help engineering teams analyze autonomous system behavior in production.

Excessive agency security risks

• ▪The OWASP Top 10 for Large Language Model Applications defines Excessive Agency as the risk when an LLM-based system is granted more functionality, permissions, or autonomy than required.
• ▪Excessive agency manifests as excessive functionality, excessive permissions, and excessive autonomy, which can escalate minor model errors into major operational incidents.
• ▪To mitigate excessive agency, organizations should implement the principle of least privilege, establish human-in-the-loop checkpoints for high-impact actions, and conduct periodic permission reviews.

AI agent monitoring signals

• ▪Security monitoring for AI agents includes detecting prompt injection attempts, identifying tool-call anomalies, and tracking egress and data-exfiltration channels.
• ▪AI agent monitoring signals are categorized into task outcome signals, behavior and tool-use signals, safety and security signals, and cost and performance signals.

5 sources