Plurai Raises $13M Seed to Train AI Agents Like Self-Driving Cars

Plurai has raised $13 million in seed funding to bring self-driving-car simulation techniques to AI agents, the startup announced Wednesday. Team8 led the round, with U&I, Velocity and Mercer participating.

The premise is blunt: today's agents aren't reliable enough to run unsupervised. LiveBench 2.0 shows Claude Fable 5.2 and ChatGPT-6 Astra succeeding on 82.9 and 75.6 out of 100 tasks, respectively. In a separate test run by AI infrastructure company EON in a simulated corporate environment, the best performer, Claude Opus 4.8, completed tasks at 76.8% accuracy.

Plurai's co-founders, Ilan Kadar and Elad Levi, previously worked at Nexar, which develops autonomous driving technology. Kadar worked with Waymo to simulate and evaluate its end-to-end deep learning driving software. At Plurai, they're applying the same playbook: simulations built on synthetic data drawn from a company's own digital systems, designed to prepare agent models for difficult edge cases.

"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," Kadar told TechCrunch. "Since we saw the same pattern, the same pitch was that we want to build this simulation layer."

Why do agents fail in production?

Like end-to-end driving models, AI agents are probabilistic systems operating in unpredictable environments. The failure modes are concrete. Agents tasked with booking travel can ignore company spending policies when pressed by customers. Agents that answer reliably when querying a single database tend to gum up when combining information from multiple data sets. Long customer-service conversations can overwhelm an agent with too much context and lead it down wrong paths.

Plurai's fix starts with a digital twin of a company's electronic systems, including the knowledge base and tools the agent will access in production. The company then maps all possible agent flows and uses that graph to prepare the model with evaluations and fixes for potential edge cases. The platform runs alongside existing agentic tools including Claude Code, Codex, Copilot and Cursor.

Where does the real value sit?

Kadar says the bigger prize is post-training small, open-weight models tailored to individual customers, using the data Plurai gathers. Tests run in cooperation with Nvidia and Nebius showed Plurai's models outperforming 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," Kadar said.

Reah Miyara, a Google DeepMind researcher who backed Plurai as an angel investor, said the company's edge comes from understanding failure modes in production and generating synthetic data to tune models. As businesses try to optimize AI spending, that matters, because IT leaders are finding that "frontier-capable models are out of distribution with my specific data, and the cost of these things isn't sustainable in the long term," Miyara said.

Does the robotics analogy hold?

Plurai is betting that patterns from physical robots carry over to digital agents. Powerful general AI models are giving robots impressive capabilities, but those robots still aren't reliable enough for wide deployment. Meanwhile, robots running task-specific AI models are already in the field, serving customers and collecting valuable real-world data.

Kadar's argument is that general intelligence won't substitute for customer-specific tuning. "My view is that no matter how smart OpenAI and Anthropic model will be — and they will be much smarter — they will not know your business and your policy," he said.

With $13 million in the bank, Plurai is now positioned to test whether simulation-driven training can close the reliability gap that benchmark scores keep exposing — and whether cost-conscious enterprises will buy smaller, tuned models over frontier ones.