AI agent reliability startup Lemma has raised $2.3 million in pre-seed funding to construct monitoring infrastructure designed to catch a very tough class of drawback: AI brokers that seem to have accomplished a job efficiently whereas quietly producing the incorrect outcome.
The spherical consists of participation from Matrix, Y Combinator, Liquid 2 Ventures, Vermilion Cliffs Ventures, Irregular Expressions, Cervin Ventures, Comma Capital, Place Ventures, and Eight Capital, alongside angel buyers and operators from OpenAI, xAI, Meta, and DoorDash.
Based by Jerry Zhang and Cole Gawin, Lemma was a part of Y Combinator’s Fall 2025 batch and focuses on manufacturing monitoring for AI brokers. The corporate says its platform has now processed a couple of million agent traces as engineering groups more and more search for methods to grasp how autonomous programs behave after deployment.
The Rising Downside of AI Brokers That Fail Silently
Conventional software program monitoring is basically designed round express failure alerts. An utility crashes, a request returns an error code, latency spikes, or an infrastructure part turns into unavailable.
AI brokers introduce a special drawback.
An agent can efficiently execute each technical step in a workflow and nonetheless misunderstand what the person needed, name the incorrect device, use incorrect data, develop into caught in an unproductive loop, or return a believable however incorrect reply. From the attitude of typical monitoring infrastructure, the request might look completely wholesome.
Lemma describes these as semantic failures. Examples embrace a customer support agent citing the incorrect refund coverage, an auditing agent producing an outdated report, or an agent calling an exterior system utilizing data it invented. These are frequent AI agent failure factors.
That distinction turns into more and more vital as brokers transfer past conversational interfaces and start executing longer, multi-step workflows the place language fashions work together with databases, utility programming interfaces (APIs), retrieval programs, and different software program instruments.
A failure someplace in that chain might not produce an exception. The agent might merely proceed.
How Lemma Displays AI Brokers in Manufacturing
Lemma is constructing an observability layer particularly round these agent execution paths.
Its tracing system turns every agent execution right into a structured hint containing the underlying giant language mannequin calls, device invocations, inputs, outputs, timing information, retrieval steps, and errors generated all through the workflow. Engineering groups can then look at a complete execution tree slightly than wanting solely on the agent’s last response.
However tracing is barely a part of the strategy.
Lemma analyzes manufacturing traces in opposition to an agent’s directions and teams recurring issues into points, serving to groups establish failure patterns which may in any other case stay buried throughout 1000’s of particular person interactions. The platform also can prioritize points and ship alerts by way of Slack when probably vital issues seem.
The target is to reply a tougher query than whether or not the software program ran efficiently: Did the agent really accomplish what it was supposed to perform?
That may be a vital shift in how observability might must work for agentic software program.
Turning Manufacturing Failures Into Agent Enhancements
Lemma can also be attempting to shorten the gap between discovering an issue and fixing it.
As soon as the platform identifies a recurring failure, it analyzes the encircling traces and context to find out a possible root trigger. From there, it may suggest modifications to prompts, utility logic, or agent workflows slightly than requiring engineers to manually reconstruct each problematic interplay.
The corporate is extending that workflow into growth environments by way of a Mannequin Context Protocol (MCP) server. Builders can question Lemma’s traces from instruments together with Cursor, Claude Desktop, and Claude Code, permitting the debugging course of to occur nearer to the place the underlying agent is being developed.
After a repair is deployed, Lemma can flip the manufacturing failure into a web based analysis and monitor for its recurrence. This creates a suggestions loop wherein beforehand unseen real-world failures develop into future checks slightly than remaining remoted incidents.
This strategy pushes Lemma considerably past typical observability. The longer-term objective is infrastructure that helps brokers study systematically from manufacturing failures slightly than relying fully on engineers to find, reproduce, and manually patch each edge case.
A Downside the Founders Encountered Firsthand
Zhang and Gawin met as freshmen on the College of Southern California and later labored on AI programs at separate AI-native startups. Earlier than founding Lemma, they labored at Tandem, which applies AI in healthcare, and ChipStack, which develops AI brokers for chip design.
These experiences helped expose them to the problem of taking brokers from managed growth environments into manufacturing.
“Cole and I began Lemma as a result of we skilled the ache of constructing AI brokers firsthand,” Zhang mentioned. “We stored working into the identical drawback: brokers would seem to work, however the outcomes weren’t dependable sufficient in manufacturing.”
The founders argue that bettering underlying basis fashions alone is not going to get rid of this drawback. Actual-world agent conduct additionally is dependent upon prompts, utility logic, instruments, integrations, retrieval programs, person conduct, and the more and more sophisticated chains connecting them.
Lemma’s personal engineering thesis is that offline evaluations battle to breed the unpredictable circumstances brokers encounter after deployment, making manufacturing information an vital supply for understanding the place programs really break down.
The Broader Problem of Monitoring AI Brokers in Manufacturing
The brand new funding will help additional growth of Lemma’s monitoring and failure-detection instruments, with an preliminary concentrate on startups already working AI brokers in manufacturing.
The corporate is working in an space that’s changing into extra vital as AI programs transfer from remoted demonstrations into real-world workflows. Conventional observability instruments are usually good at detecting technical issues reminiscent of downtime, latency, or failed requests, however agentic programs introduce one other layer of complexity: an utility can stay operational whereas the agent misunderstands a job, chooses the incorrect device, or produces an incorrect outcome.
That distinction is prone to develop into extra vital as brokers are used throughout buyer help, monetary evaluation, healthcare administration, software program growth, and analysis. In these environments, measuring whether or not an agent accomplished a workflow might matter lower than figuring out whether or not it accomplished the workflow accurately.
For Lemma, the chance due to this fact is dependent upon whether or not monitoring semantic failures turns into a typical a part of working AI brokers in manufacturing. The $2.3 million pre-seed spherical provides the corporate further capital to check that thesis as organizations deploy brokers throughout more and more advanced workflows.
What Higher Agent Monitoring May Imply for AI
As AI brokers tackle extra advanced and autonomous work, conventional monitoring might not be sufficient. Future programs might want to assess not solely whether or not an agent accomplished a job, however whether or not it understood the target, used the appropriate instruments, and produced the right end result.
Instruments like Lemma might additionally create tighter suggestions loops between manufacturing and growth, turning real-world failures into new checks and enhancements. Over time, this might make agent observability a typical a part of the AI infrastructure stack, significantly in high-stakes environments the place reliability and accountability matter most.
