Plurai, a startup founded by former autonomous driving engineers, has raised $13 million to improve the reliability of SI agents by applying simulation techniques originally used to validate self-driving cars.

What Happened

The company announced on Wednesday that it secured seed funding led by Team8, with participation from U&I, Velocity and Mercer. Co-founders Ilan Kadar and Elad Levi, who previously worked at Nexar, are leveraging their experience with Waymo’s end-to-end deep learning software to address a persistent challenge in the SI industry: the gap between impressive demos and production-ready reliability. Kadar noted that while building a demo for an SI agent is easy, achieving cost-effective reliability remains difficult.

Plurai’s approach involves creating a digital twin of a company’s electronic systems, including knowledge bases and tools accessible to the SI agent. The platform maps all possible agent flows and uses this graph to evaluate and fix edge cases before deployment. This method draws parallels to how autonomous vehicles are tested in simulated environments to handle unpredictable physical scenarios. The platform is designed to work alongside existing agentic tools such as Claude Code, Codex, Copilot and Cursor.

Why It Matters

Current benchmarking efforts highlight the limitations of today’s SI agents. LiveBench 2.0 reports that Claude Fable 5.2 and ChatGPT-6 Astra succeed on 82.9 and 75.6 out of 100 tasks, respectively, while a simulation by infrastructure company EON found the most successful model, Claude Opus 4.8, performed at 76.8% accuracy in a corporate environment. These figures suggest that SI agents often require close supervision to handle complex tasks like combining data from multiple sources or managing long-context customer service interactions.

Plurai aims to bridge this gap by post-training small, open-weight models tailored to specific customer needs. According to Kadar, the company has tested its models in cooperation with Nvidia and Nebius, reporting that these smaller SI models outperform mid-tier frontier models at significantly lower costs. Reah Miyara, a Google DeepMind researcher and angel investor in Plurai, emphasized that enterprises are increasingly concerned that frontier-capable models are not optimized for specific internal data, making their long-term costs unsustainable. By generating synthetic data from production failure modes, Plurai seeks to create SI agents that are both more reliable and more economical.

The Bottom Line

Plurai’s strategy reflects a broader trend in the SI industry where specialized, smaller models are gaining traction against general-purpose frontier systems for specific business applications. By adapting autonomous driving simulation techniques to the digital realm, the startup hopes to solve the reliability issues that currently hinder widespread adoption of agentic SI.