Plurai Raises $13 Million to Test AI Agents in Simulated Worlds
Plurai has raised $13 million in seed funding, led by Team8, to build simulation and post-training tools aimed at the reliability of AI agents. U&I, Velocity and Mercer also participated in the round, according to TechCrunch, which reported the raise on November 11, 2026.
The company was co-founded by Ilan Kadar and Elad Levi, who previously worked at Nexar, the autonomous driving technology developer. Kadar also worked with Waymo to simulate and evaluate the company's end-to-end deep learning driving software. Plurai is applying that simulation-first approach to agentic systems, meaning AI systems in which models plan and take actions through tools. Google DeepMind researcher Reah Miyara has backed the company as an angel investor.
The core product is a digital twin, a software replica, of a company's electronic systems. The twin includes the knowledge base and the tools an agent will access in production, so the agent can be exercised against a synthetic environment before it touches live data. Plurai's website describes the platform as a recursive self-improvement layer for AI agents, a loop in which performance feeds back into tuning, that combines agentic world simulation with model post-training. Post-training is additional training applied after a base model is built. Its Simulation product maps an agent as built and mocks, or builds stand-ins for, the tools, sub-agents and knowledge the agent uses to run real tasks.
Plurai's platform is designed to work alongside Claude Code, Codex, Copilot and Cursor, the coding agents and assistants that many engineering teams already use. The company plans to feed the data it gathers into post-training small, open-weight models, meaning models whose trained weights are publicly released, tailored to individual customer needs. That is a stated plan.
Plurai's website makes several performance claims. It states that its platform delivers 98% inference cost savings, 85% lower latency and 82% fewer production failures. For its Post-training product, it claims more than 44x lower cost than Claude Haiku-5.5, more than 10% higher accuracy and inference latency under 100ms.
The company published a research blog post on November 11, 2026, titled "Production bugs your agent won't repeat, and how we pin them down," written by Reut Vilek. It has scheduled a webinar, "Beyond the leaderboard: AI agents that learn in simulated worlds," for December 9, 2026 at 8:00 AM PT, featuring its founders.
In my view, the Waymo lineage is the more instructive part of this story. Self-driving teams learned early that a model's field behavior depends heavily on the scenarios it was evaluated against, and agent builders are now hitting the same problem with tool calls, retrieval and multi-step workflows. The open question is whether a synthetic twin reproduces the failure modes that production actually generates. Plurai's blog post on production bugs is one place where that claim will be tested in public.
The broader context here is a shift in where agent quality gets worked out. Much of the discussion has centered on model benchmarks. Plurai moves the evaluation step into a replica of the customer's own stack, and pairs that with smaller, cheaper models trained on the results. If the cost and latency figures hold under independent measurement, the economics of running agents in production could change. Until then, those figures remain Plurai's claims, and a seed-stage company with a $13 million round still has to show that its simulated environments match the systems its customers run.
