YC startup builds cheap AI model that catches silent agent failures

The problem with silent failures

AI agents are becoming more common in software development, legal work, and customer support. But as Laminar co-founder Robert explained on Hacker News, agents can fail silently — clicking the wrong button or citing an incorrect document while every individual API call completes successfully ¹.

Many are essentially data collectors that store traces nobody has time to read.

A model trained to investigate

Laminar built flow-1, a small open-source model finetuned and then post-trained with reinforcement learning on thousands of synthetic investigations of realistic agent workflows ¹. The model was designed to act as a coding agent, treating each agent trace like a repository where every LLM call or tool interaction is a file.

On Laminar's benchmark, flow-1 matches the detection quality of GPT-6-sol while being 23 times cheaper to run ¹. This cost advantage makes it affordable to analyze every production agent trace without relying on sampling.

How it works in practice

The model runs inside a harness Laminar calls Signals, which reads compressed trace representations and greps across spans to find relevant context ¹. It can respond to general prompts beyond raw error detection — identifying user frustration, verifying checkout completion, or checking for other specific conditions.

Findings are returned as JSON in schemas users can define, called signal events. Similar events are automatically grouped into clusters so teams can spot recurring problems and drill into individual traces.

Open source and self-hosted

Laminar is releasing its platform under the Apache 2.0 license with self-hosting options available on GitHub ¹. The company makes money from its managed cloud offering. The founders say their method produced three main results: trace compression that is 20 times better, a purpose-built SQL query engine, and the ability to extract tasks automatically.

The project comes from Robert and Din, who have been building open-source observability for AI agents since 2024.

Corrections and updates

• Corrected on 23 October 2026. An earlier version said Laminar was applying for Y Combinator's Winter 2027 batch. The source indicates the company is associated with Y Combinator's Summer 2024 batch (YC S24), not an application for Winter 2027, so the statement was removed. An earlier version said the hidden errors were 'the most damaging because they go undetected,' but the source stated only that such errors are 'the most useful ones to catch and understand' — it did not describe them as most damaging.

Key facts

• flow-1 matches GPT-6-sol detection quality at 23x lower cost
• Model trained with reinforcement learning on synthetic agent workflows
• Laminar platform is Apache 2.0 licensed and self-hostable
• Automatically clusters similar signal events for pattern detection
• Handles traces from software, legal, and customer support agents

Sources

1. Launch HN: Laminar (YC S24) – We trained a model to catch silent agent failures | Hacker News — news.ycombinator.com