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

Plurai, a startup founded by two former Nexar executives, said Wednesday it has raised $13 million in seed funding to build a simulation layer for AI agents. The round was led by Team8, with participation from U&I, Velocity and Mercer. The company's pitch borrows directly from autonomous driving: prepare probabilistic systems for hard edge cases with simulations built from synthetic data generated off a company's own digital systems, rather than discovering those failures in production.

Co-founders Ilan Kadar and Elad Levi previously worked at Nexar, which builds autonomous driving technology, and Kadar worked with Waymo to simulate and evaluate its end-to-end deep learning driving software. Kadar said the parallel with agents was immediately obvious. "The moment that we saw the wave of AI agents, we saw the same pattern: That it is very easy to build a demo, but [hard] to get something that will be reliable and cost effective," he told TechCrunch, adding that the startup's "same pitch" is to build that simulation layer.

The reliability gap Plurai targets is documented in recent evaluations. LiveBench 2.0 shows Claude Fable 5.2 and ChatGPT-6 Astra succeeding on 82.9 and 75.6 out of 100 tasks respectively, while infrastructure company EON's simulated corporate environment found the best model it tested, Claude Opus 4.8, completing tasks at 76.8% accuracy. The failure modes are mundane but persistent: travel-booking agents ignoring spending policies, queries that work against one database but break when information is combined across data sets, and long customer service conversations whose accumulated context pushes an agent into wrong turns.

Plurai's approach is to construct a digital twin of a customer's electronic systems, including the knowledge base and tools the agent will use in production. It then maps every possible flow the agent can take and uses that graph to run evaluations and apply fixes for likely edge cases. The platform is designed to sit alongside existing agentic tooling such as Claude Code, Codex, Copilot and Cursor rather than replace it.

Kadar framed the deeper bet as post-training small, open-weight models on the data the simulation generates, tuned to individual customers. Tests conducted with Nvidia and Nebius showed the company's models beating mid-tier frontier models at lower cost. "We can take a large language model, we can take the task, decompose it into a smaller task using the simulation, train [an open-weight] model, and using the simulation, verify that the quality is higher or comparable to the frontier model, and an order or two orders of magnitude cost reduction," he said.

What to watch is whether those simulated edge cases translate into measured gains on real customer workflows, and whether the one-to-two-order-of-magnitude cost claim holds as tasks are decomposed further. The open question for builders is also architectural: whether customers will keep running generalist frontier models alongside the tooling they already use, or move core agent workloads to the tailored open-weight models Plurai is training.