Plurai Raises $13M to Make AI Agents Reliable With Simulation

Why agents still need supervision

Benchmarks show agents are not yet reliable enough to work unsupervised. LiveBench 2.0 has Claude Fable 5.2 and ChatGPT-6 Astra completing 82.9 and 75.6 out of 100 tasks respectively, while a test by AI infrastructure company EON in a simulated corporate environment found the best model, Claude Opus 4.8, reached 76.8% accuracy.

Kadar describes agents as probabilistic systems operating in unpredictable environments, much like driving models. Travel-booking agents can ignore spending policies under customer pressure, combining multiple data sets tends to trip them up, and long customer-service conversations can overload them with context.

Plurai builds a digital twin of a company's electronic systems, including the knowledge base and tools an agent uses in production. It maps all possible agent flows and uses that graph to prepare the model with evaluations and fixes. The platform works alongside Claude Code, Codex, Copilot and Cursor.

Smaller models, lower costs

Kadar says the bigger promise is using the collected data to post-train small, open-weight models for individual customers. In tests with Nvidia and Nebius, Plurai's models outperformed mid-tier frontier models at lower cost. He claims quality equal to or better than frontier models with a cost reduction of one to two orders of magnitude.

Google DeepMind researcher Reah Miyara, an angel investor in Plurai, says IT leaders find frontier models poorly matched to their own data and too costly long term. Kadar adds that however smart models from OpenAI and Anthropic become, "they will not know your business and your policy."