Plurai raises USD $13 million for AI agent testing platform

Plurai has raised USD $13 million in seed funding and opened its AI agent testing platform to all developers. Team8 led the round.

The New York-based company says its software is designed to identify failures in AI agents before they are used by customers and to monitor them in production. According to Plurai, design partners have used the platform to cut development cycles from months to hours, reduce failures reaching users by 82% and lower running costs by 97%.

Plurai's system creates what it describes as a digital twin of an agent and its operating environment using a customer's knowledge base, tools and policies. It then runs large volumes of simulated and adversarial user interactions in a sandbox to uncover problems such as policy breaches, hallucinations, incorrect actions, broken tool calls, failed tasks and data leaks.

The same simulation process is also used to improve the agent after testing. It generates fixes for failures and uses simulation data to create small language models tailored to specific tasks. These can either run the agent itself or act as evaluators and guardrails during live use.

The approach targets a problem many companies face as they move AI agents into customer-facing and internal workflows. Plurai cited research from VentureBeat Pulse Research that found half of enterprises had deployed an AI agent or large language model feature that passed internal evaluation but still caused a customer-facing failure in production.

Backers and advisers

Investors in the seed round included U&I, Velocity and Mercer, alongside more than 20 angel backers and industry figures, including CyberArk Founder Udi Mokady. Microsoft 365 Copilot Chief Technology Officer Dr Eyal Carmi has also joined the company's advisory board.

Plurai also said it had worked with NVIDIA and Nebius on a project for a global enterprise customer, where models it post-trained outperformed frontier models from OpenAI and Anthropic on accuracy at about 40 times and 44 times lower cost, respectively. The company did not identify the customer.

The platform, previously limited to selected partners and customers, is now available more broadly with a free tier. Developers can start with tools including Claude Code, Codex and GitHub Copilot, while teams can also test and improve agents built on enterprise platforms such as Salesforce Agentforce.

Founders' background

Plurai was founded by Dr Ilan Kadar and Dr Elad Levi, who previously worked at Nexar on AI systems for autonomous driving. Kadar was Vice President of AI, and Levi was an AI Architect. Both were involved in building systems used in safety-critical settings, where simulation is commonly used before real-world deployment.

The company says its technology also draws on its own research, including a method called BARRED for generating synthetic training data for custom small language models. It also cited work on serving hundreds of tailored models in real time on a single graphics processing unit.

"AI systems fail when they hit challenges down the road which no one predicted," said Dr. Kadar, Chief Executive Officer of Plurai. "Waymo built trust by testing and training across more than 20 billion simulated miles, roughly 100 times its real-world driving. Plurai brings that same approach to AI agents: continuously challenging them with the hardest situations and rarest events before they reach users, learning from failures, and improving before those failures happen in the real world."

Team8 said Plurai is addressing a gap in the software stack as businesses look for ways to test AI systems before wider deployment.

"As agents move into core enterprise workflows, the missing layer is infrastructure that builds trust in them before they reach customers and keeps them improving in production," said Aviad Harell, Managing Partner at Team8. "Simulation will become as fundamental to the agent stack as CI/CD is to software. Ilan and Elad bring deep AI expertise and more than a decade of experience bringing AI into high-stakes production systems, and they are the team to build that layer."

Other backers and advisers pointed to the constraints companies face when testing AI tools on real user data, particularly in regulated sectors and in systems with privacy and security risks.

"AI is probabilistic. Enterprises are not," said Mokady. "The founders have spent more than a decade building AI for mission-critical environments like autonomous driving, where systems do not get second chances. That experience shows in how they approach agents, and it is what separates AI that can be deployed from AI that can be trusted."

"For enterprises, evaluating AI agents on real customer data is often blocked by privacy, security and compliance constraints, which makes simulation mission-critical," said Carmi. "What sets Plurai apart is how faithfully it recreates each company's world, from its tools and policies to the users its agents actually face. And it does not stop at finding failures. It turns what it learns into fixes and production safeguards, so teams can deploy agents with confidence."

"Frontier models are incredibly capable, but using them to continuously evaluate agent behavior in production can quickly become expensive and impractical," said Reah Miyara, Senior Director of Google Models at Google Cloud. "What I find compelling about Plurai is the idea of using smaller, tailored models to make evaluation continuous, not just something you do on a small sample of traffic."