Despite the growing enthusiasm around AI agents, the technology still falls well short of the reliability that many vendors imply. Independent benchmarking efforts continue to expose significant gaps between what these systems promise and what they actually deliver in practice. LiveBench 2.0 recorded success rates of 82.9 and 75.6 out of 100 tasks for two leading models, while a separate evaluation conducted by AI infrastructure firm EON placed the top-performing model, Claude Opus 4.8, at just 76.8% task accuracy when tested inside a simulated corporate environment. These numbers point to a technology that, while impressive, still requires close human oversight to function dependably in real-world business settings.

Into that gap steps Plurai, a startup that believes the methods used to develop self-driving cars hold the key to making AI agents more capable and consistent. The company is building simulation environments based on synthetic data derived from a client's own digital infrastructure - a technique designed to expose and address the edge cases that cause AI agents to fail when deployed at scale.

Founders With Roots in Autonomous Driving

Plurai was founded by Ilan Kadar and Elad Levi, both of whom previously worked at Nexar, a company focused on autonomous driving technology. Kadar's background includes work with Waymo, where he was involved in simulating and evaluating end-to-end deep learning software used to power self-driving vehicles. That experience gave him a direct view of how simulation-based training can transform an unreliable prototype into a system robust enough for real-world deployment.

When AI agents began gaining traction across the industry, Kadar and Levi recognized a familiar dynamic. "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 explained. "Since we saw the same pattern, the same pitch was that we want to build this simulation layer."

On Wednesday, Plurai announced it had secured $13 million in seed funding to advance that vision. The round was led by Team8, with additional investment from U&I, Velocity, and Mercer.

Why AI Agents Break Down in the Real World

Much like the deep learning models trained to navigate roads, AI agents are probabilistic systems operating in environments full of unpredictability. The specific ways they fail tend to follow recognizable patterns. A travel-booking agent, when pressured by a user, may quietly disregard a company's spending policy. An agent that performs well when querying a single database can become unreliable when asked to pull and reconcile information from multiple data sources. Customer service agents handling extended conversations can become overwhelmed by accumulated context, leading to errors that compound over time.

These are not simply bugs to be patched. They reflect a deeper challenge: AI agents encounter an enormous range of situations in production, and no amount of pre-deployment testing can anticipate every scenario without a systematic approach to identifying and addressing failure modes before they reach end users.

Building a Digital Twin for Agent Training

Plurai's response to this challenge is to construct a digital twin of a client's electronic systems - including the knowledge bases, databases, and tools that an agent will interact with once deployed. From that foundation, the platform maps every possible flow the agent might take through those systems, generating a comprehensive graph of interactions. That graph then drives a process of evaluation and targeted remediation, identifying where the agent is likely to stumble and preparing it to handle those situations correctly.

The platform is designed to work alongside existing agentic development tools, including Claude Code, Codex, Copilot, and Cursor, allowing teams to integrate Plurai's simulation layer without abandoning their current workflows.

The Bigger Opportunity: Post-Training Smaller Models

While the simulation and evaluation layer addresses immediate reliability concerns, Kadar points to a more significant long-term opportunity: using the data gathered through simulation to post-train small, open-weight language models that are specifically tuned to each customer's needs and environment.

Tests conducted in partnership with Nvidia and Nebius demonstrated that Plurai's purpose-built models can outperform mid-tier frontier models while running at substantially lower cost. The process, as Kadar describes it, involves taking a large language model, decomposing a complex task into smaller components using simulation, training an open-weight model on that structured data, and then using the simulation environment to verify that the resulting model performs at a level comparable to - or better than - a frontier model, but at a fraction of the price.

"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.

Investor Perspective: Failure Modes and Cost Pressure

Reah Miyara, a researcher at Google DeepMind who participated in Plurai's round as an angel investor, framed the company's advantage in terms of its ability to understand where agents fail in production and to generate synthetic data that addresses those failures through targeted model tuning.

Miyara noted that as businesses look to manage and reduce their AI spending, the limitations of large frontier models are becoming increasingly apparent to IT decision-makers. The core tension is that general-purpose frontier models are not trained on a company's specific data and policies, and the cost of running them at scale is proving difficult to justify over the long term. "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, capturing the concern driving many enterprise buyers toward alternatives.

A Parallel With Physical Robotics

Plurai's broader thesis draws on what is already playing out in the physical world. Powerful general-purpose AI models are enabling robots to perform impressive and varied tasks, but those systems are not yet reliable or consistent enough to be deployed widely in demanding operational settings. By contrast, robots trained on AI models built for specific, well-defined tasks are already working in the field, serving customers and accumulating real-world data that further refines their performance.

Plurai believes the same dynamic will shape how AI agents evolve in digital environments. General models from leading labs will continue to grow more capable, but they will not automatically become experts in any given company's internal systems, workflows, or policies. That gap, the company argues, is where purpose-built, simulation-trained agents will consistently outperform their generalist counterparts.

"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," Kadar said. It is a bet that domain-specific knowledge, built through rigorous simulation, will remain a durable competitive advantage even as the underlying capabilities of frontier AI models continue to advance.