Lemma raises $2.3M pre-seed to catch AI agents that fail silently
What's the deal? LemmaDealroom has a profile for this one. Try Dealroom →, an AI agent reliability startup, has raised $2.3 million in pre-seed funding to build monitoring infrastructure that catches agents which appear to complete a task while quietly producing the wrong result. MatrixDealroom has a profile for this one. Try Dealroom →, Y Combinator, Liquid 2 Ventures, Vermilion Cliffs VenturesDealroom has a profile for this one. Try Dealroom →, Irregular ExpressionsDealroom has a profile for this one. Try Dealroom →, Cervin VenturesDealroom has a profile for this one. Try Dealroom →, Comma CapitalDealroom has a profile for this one. Try Dealroom →, Position VenturesDealroom has a profile for this one. Try Dealroom →, and Eight CapitalDealroom has a profile for this one. Try Dealroom → took part. Angel investors and operators from OpenAI, xAI, Meta, and DoorDash also joined.
What’s the endgame? Founded by Jerry Zhang and Cole Gawin, Lemma builds production monitoring for AI agents. Its platform has processed more than one million agent traces.
What could go wrong? Traditional software monitoring watches for explicit failures — crashes, error codes, latency spikes. AI agents introduce what Lemma calls semantic failures: an agent can execute every technical step correctly yet misunderstand the user, call the wrong tool, or return a plausible but incorrect answer.
Examples include a customer service agent citing the wrong refund policy or an auditing agent generating an outdated report. To conventional monitoring, the request looks healthy.
How it works: Lemma turns each agent execution into a structured trace covering the underlying model calls, tool invocations, inputs, outputs, and retrieval steps. It groups recurring problems into issues and sends alerts through Slack.
The platform then analyses the surrounding traces to find a likely root cause and can propose changes to prompts, logic, or workflows. Developers can query its traces from tools including CursorDealroom has a profile for this one. Try Dealroom →, Claude DesktopDealroom has a profile for this one. Try Dealroom →, and Claude Code via a Model Context Protocol (MCP) server.
Why now? The problem grows as agents move beyond chat into longer, multi-step workflows where models interact with databases, APIs, and other software. A failure in that chain may not throw an exception — the agent simply continues.
The signal: Lemma is betting that observability for agentic software needs to answer a harder question than whether the code ran: did the agent do what it was supposed to do?
Read more: Unite.AI
Image credit: Generated with Gemini
