Plurai Wants to Make AI Agents Reliable Using the Same Tricks as Self-Driving Cars
The challenge of making AI agents reliable in production is the same one that held back the self-driving car industry for years: it's easy to build a demo in the lab, but reality reveals itself to be infinitely more complex when the system encounters edge cases nobody predicted. Plurai, an Israeli startup that just emerged from stealth with $13 million in seed funding, wants to solve exactly that problem by bringing to AI agents the same simulation arsenal that finally put self-driving cars on the road.
Co-founders Ilan Kadar and Elad Levi know that pain intimately. At their previous company, Nexar—an autonomous driving technology firm that served Waymo and others—they saw firsthand how the autonomous vehicle industry earned public trust: not with bigger models, but with massive simulations based on the company's own real-world data. Waymo would accumulate more than 20 billion simulated miles, roughly 100 times the actual distance its vehicles drove. The rarest events were resolved first in simulation.
Why agents fail
The current numbers are revealing. The LiveBench 2.0 benchmark shows that Claude Fable 5.2 scored 82.9 and ChatGPT-6 Astra scored 75.6 out of 100 evaluation tasks. Meanwhile, an effort by AI infrastructure company EON to test agents in a simulated corporate environment found the most successful model tested, Claude Opus 4.8, with only 76.8% accuracy. In other words: even the most advanced models fail on a quarter of tasks.
"When we saw the wave of AI agents, we recognized the same pattern," Kadar told TechCrunch. "It's very easy to build a demo, but very hard to get something reliable and cost-effective. We saw the same pattern and the same pitch, which is why we wanted to build this simulation layer."
The problem is structural. AI agents, like end-to-end models trained to drive, are probabilistic systems operating in unpredictable environments. When pressed by real users, agents tasked with booking travel can ignore spending policies; customer service assistants can hallucinate answers that sound convincing but are false; IT bots can execute irreversible actions based on ambiguous instructions.
Plurai doesn't try to solve this with bigger models. Instead, it creates a "digital twin" of the agent and the environment in which it will operate, built from the company's own knowledge base, tools, and policies. The system then simulates thousands of realistic scenarios—including adversarial users—and exposes policy violations, hallucinations, broken tool calls, and data leaks before any customer encounters them.
Post-training: the real promise
Finding failures is only half the battle. Plurai's greater promise lies in what it does with that data: using what it learns to post-train small, open-weight models tailored to each customer's specific needs.
In tests done in partnership with Nvidia and Nebius, the company claims a Plurai post-trained model outperformed frontier models from OpenAI and Anthropic in accuracy at costs 40x and 44x lower, respectively. A customer case study (provided in August 2026) reported inference time reduced from $2.87 to $0.15 per 1,000 calls and inference cost from $2.87 to $0.076 per 1,000 calls.
The technology behind this includes a synthetic data generation methodology based on intent calibration—an approach Plurai developed internally that was accepted for presentation at ICML 2026. The technique, described in a preprint paper submitted in April 2026 by Arnon Mazza and Levi, allows developers to turn open AI models into specialized models for specific tasks in about 20 minutes.
"My view is that no matter how smart OpenAI and Anthropic models will become—and they will become much smarter—they won't know your business and your policies," Kadar said. "Simulation is what will close that gap."
What to watch
Plurai is not the first to address AI agent evaluation—companies like EON and others already exist in the space. The difference lies in the team's origin, with decades of experience in autonomous vehicle simulation, and in the post-training strategy: rather than just measuring failures, Plurai uses the generated data to improve the actual models.
Gartner named Plurai a Representative Vendor in its 2026 Market Guide for AI Evaluation and Observability Platform, and the company claims to work with Fortune 500 clients in highly regulated industries. As of today, the platform is available in a free tier for any agent builder, integrable directly with Claude Code, Codex, GitHub Copilot, and Cursor.
The AI agent market is moving at a speed that makes production reliability a matter of survival, not just quality. The question is whether simulation—the same technique that took the self-driving car industry decades to mature—can accelerate the same path for agents, or whether the complexity of AI agents exceeds any analogy with autonomous driving.
Sources: TechCrunch, Plurai, Team8
✓ Independent sources cross-checked and verified before publishing
