Plurai opens a simulation platform for enterprise AI agents
US startup Plurai has opened its platform for testing AI agents in simulated company environments. The system recreates internal tools, knowledge and policies to expose failures before customer use and generate task-specific training data. Its founders bring experience from autonomous-driving simulation, applying that approach to software agents handling business workflows and longer customer conversations.
Artificial Intelligence··Night
A company’s tools become a simulated environment
US startup Plurai opened its AI-agent simulation platform on 11 November after working with a limited group of partners. The service is intended to find failures before agents reach customers and continue evaluating them during production. It recreates an organization’s knowledge base, software tools and policies in a digital twin: a simulated version of the environment in which the agent is expected to operate.[1], [2]
Possible agent workflows are mapped into a graph that supports testing and proposed corrections. TechCrunch describes examples including a travel-booking agent overlooking spending rules, difficulty combining multiple datasets and long customer-service conversations overwhelming the agent’s context. These are the kinds of task-specific situations Plurai seeks to reproduce using the customer’s own systems rather than a generic set of questions.[1], [2]
Simulated failures feed task-specific training
Plurai says it runs realistic and adversarial simulated users against agents in an isolated environment. The tests target incorrect actions, failed tool calls, invented answers, sensitive-data leaks and departures from company policy. The resulting examples are then used to generate fixes and train smaller language models for the agent’s particular work. The same simulated environment is used to evaluate the proposed changes.[2]
Co-founder Ilan Kadar describes breaking a task into smaller operations and using simulation data to post-train open-weight models. In the company’s design, those models can run an agent or act as continuous evaluators during live use. Failures found in production feed back into the simulation process, allowing the test environment to incorporate problems encountered after the initial deployment.[1], [2]
Autonomous-driving experience informs the platform
Kadar and fellow founder Elad Levi previously worked at Nexar, an autonomous-driving technology company. Kadar also worked with Waymo on simulation and evaluation of end-to-end driving software. They describe a similar challenge in business agents: a demonstration can work while uncommon situations expose failures in real operation. The platform applies simulation methods to software tools, internal information and company rules.[1], [2]
Developers can use Plurai alongside agent-building tools including Claude Code, Codex, GitHub Copilot and Cursor. The company also describes testing agents built on enterprise platforms such as Salesforce Agentforce.[1], [2]
