Plurai gets $13 million to test AI agents in simulations
Plurai tells that small models do the work of large models at lower cost, but it gave no numbers.
Plurai, a startup that tests AI agents in simulations before use, got $13 million in seed funding from four investors. Team8 led the funding, and Plurai gave no test numbers for its claims that the cost is lower. It makes a digital twin, a copy of the systems of a company, for agents to try hard tasks.
How it works
Plurai copies the knowledge base and tools that an agent uses. It makes a map of all the flows of an agent, and tests each one with hard tasks.
Ilan Kadar, a founder, tells that Plurai will use the data to train small open weight models for each customer. Anyone can get and use open weight models.
The proof
Agents are not always reliable. On LiveBench 2.0, two models do 82.9 and 75.6 of 100 tasks. In a test by EON, the best model did 76.8% of the tasks correctly.
The firm tells that tests with Nvidia and Nebius show more quality than middle tier frontier models, at lower cost. The source gives no figures.
What is not known
Kadar tells that the cost cut is one or two orders of magnitude. The source gives no proof for this.
Reah Miyara of Google DeepMind is an angel investor in the firm.
Insight: Testing agents becomes a market
In October, two other companies released tools that check agents and connect them to owners. Plurai tests agents in simulations and then trains small models. This reads as a move to control the error data of a company and pull spend from large models. Kadar tells that large models will not know your business.
Sources
Posted
