AI agents often fall short of the capabilities their makers advertise: benchmarks and independent tests show that agents assigned complex tasks are not yet uniformly reliable without human oversight. For example, LiveBench 2.0 reports Claude Fable 5.2 and ChatGPT-6 Astra succeeded on 82.9 and 75.6 out of 100 tasks, respectively, while a corporate simulation by AI infrastructure firm EON found the best-tested model, Claude Opus 4.8, reached 76.8% accuracy.
Plurai, a startup, aims to raise these reliability numbers by applying the same simulation-driven approach used in autonomous driving: creating synthetic data and simulated environments derived from a company’s own systems so models can be trained and evaluated on difficult edge cases before deployment.
Founders and background
Plurai was founded by Ilan Kadar and Elad Levi, both of whom previously worked at Nexar, a company that develops autonomous driving technology. Kadar has experience simulating and evaluating end-to-end deep learning driving software in cooperation with Waymo. As Kadar told TechCrunch, building demos is easy, but producing reliable and cost-effective systems is hard — which motivated the team to build a simulation layer for agents.
How the platform works: digital twin and flows
Plurai creates a digital twin of a customer’s electronic systems, including the knowledge base and the tools an agent would access in production. It maps all possible agent flows on that model and uses the resulting graph to run evaluations and fixes aimed at potential edge cases.
The platform is designed to work alongside existing agentic tools such as Claude Code, Codex, Copilot and Cursor. It addresses typical failure modes: agents are probabilistic and can misapply policies when booking travel, give accurate answers from a single database but fail to integrate multiple sources, or be overwhelmed by long customer conversations and lose the correct context.
Post-training, cost savings and tests
Kadar says Plurai’s main promise is to use the data gathered by simulation to post-train small, open-weight models customized for individual customers. Tests conducted in cooperation with Nvidia and Nebius showed these models outperforming mid-tier frontier models while operating at lower cost.
According to Kadar, the process involves taking a large language model and a task, decomposing the task into smaller subtasks using the simulation, training an open-weight model, and then verifying via the simulation that the quality is higher or comparable to the frontier model — often with one or two orders of magnitude lower cost.
Funding and backers
On Wednesday the startup announced it had raised $13 million in seed funding to advance the product. The round was led by Team8, with participation from U&I, Velocity and Mercer. Reah Miyara, a Google DeepMind researcher who invested as an angel in Plurai, highlighted that the company’s competitive edge is understanding production failure modes and generating synthetic data to tune models.
Miyara noted that as enterprises try to optimize AI spend, an advantage emerges from recognizing that frontier-capable models are often out of distribution with a company’s specific data and that the ongoing cost of using those models may not be sustainable.
Why this matters
Plurai relies on an analogy between AI agents in digital environments and autonomous systems in the physical world: simulation and synthetic data can reveal and remediate failures before real-world deployment. While ever-more-powerful general models will continue to advance, Kadar argues they will not automatically know a company’s internal policies and business context, which is why company-specific simulation and post-training remain important for reliable, cost-effective agent deployment.
