Plurai wants to help AI agents work better using the same approach that helped get self-driving cars to roads

Summary

• AI agents currently lack reliability for unsupervised complex work, with benchmarks showing Claude Fable 5.2 and ChatGPT-6 Astra succeeding on 82.9 and 75.6 out of 100 tasks respectively, while Claude Opus 4.8 achieved 76.8% accuracy in simulated corporate environments.
• Startup Plurai aims to improve agent reliability using simulation techniques based on synthetic data derived from a company's own digital systems to handle difficult edge cases.
• Co-founders Ilan Kadar and Elad Levi previously worked at Nexar, with Kadar having experience simulating and evaluating Waymo's deep learning driving software.
• Plurai raised $13 million in seed funding led by Team8, with participation from U&I, Velocity, and Mercer.
• AI agents function as probabilistic systems that can fail in unpredictable environments, such as ignoring spending policies during travel bookings or struggling to combine information from multiple datasets.
• The Plurai platform creates a digital twin of a company's electronic systems, maps all possible agent flows, and uses the resulting graph to prepare models with evaluations and fixes for edge cases.
• The platform operates alongside existing agentic tools including Claude Code, Codex, Copilot, and Cursor.
• Plurai plans to use gathered data to post-train small, open-weight models tailored to individual customer needs, with tests involving Nvidia and Nebius showing these models outperform mid-tier frontier models at lower costs.
• Angel investor Reah Miyara notes that Plurai's advantage lies in understanding production failure modes and generating synthetic data to tune models for businesses seeking to optimize AI expenses.
• Founders believe that while general AI models will become smarter, they will not inherently know specific business policies, making task-specific models necessary for reliable deployment.