Plurai, a New York startup founded by former autonomous-driving researchers, has opened Plurai Simulation to every agent builder, with a free starting tier, and has announced $13 million in seed funding led by Team8. The company says the product builds a digital twin of an agent and its world, then runs realistic and adversarial simulated users against it to catch broken tool calls, policy violations, hallucinations and data leaks before customers see them. It also claims the same simulation data can be used to post-train smaller models tailored to that agent.
Quick Take
- What launched: General availability of Plurai Simulation on November 11, 2026, with a free tier, after a period limited to design partners.
- The funding: A $13 million seed round led by Team8, with U&I, Velocity and Mercer, plus more than 20 angels, including CyberArk founder Udi Mokady. Dr. Eyal Carmi, CTO of Microsoft 365 Copilot, joined the advisory board.
- The pitch: Simulate an agent against a long tail of users, explain and generate fixes for failures, then post-train a smaller model and repeat in production.
- Company-reported results: Plurai says design partners cut development cycles from months to hours, reduced failures reaching users by 82%, and lowered the cost of running reliable agents by 97%. In joint work with NVIDIA and Nebius, it says post-trained models beat unnamed OpenAI and Anthropic frontier models on accuracy at 40x and 44x lower cost. Those figures have not been independently audited in the announcement.
- How you start: The company says developers can begin from Claude Code, Codex and GitHub Copilot, and can bring agents built on Salesforce Agentforce.
In plain English
A startup is selling a practice arena for AI agents, the way self-driving labs use simulated roads before a car meets traffic. You point it at your agent, your tools and your rules. It invents lots of fake users, including ones trying to break the rules, and shows you where the agent fails, before a real customer is involved. The company also raised $13 million and says a basic version is free. The impressive percentages in its press materials are its own claims, not an independent test.
What happened
In a November 11 blog post, co-founder and CEO Dr. Ilan Kadar (with co-founder Dr. Elad Levi) wrote that Plurai was built to be “the recursive self-improvement layer for AI agents, starting with simulation.” The product, as described, creates a digital twin “from the company’s own knowledge base, tools and policies,” fills a sandbox with simulated users, and flags “policy violations, hallucinations, incorrect actions, broken tool calls, failed tasks and sensitive data leaks.” For each failure it “explains why it happened and generates the fix.” The same data, the company says, is then used to post-train models “tailored to the agent,” and production failures feed back into the simulator.
A company press release on the same day repeated the $13 million figure, the investor list and the partner metrics. TechCrunch reported the round and described the same digital-twin approach, and noted that the platform is meant to work alongside tools such as Claude Code, Codex, Copilot and Cursor. Team8’s investment note restates the design-partner claims (cycles from months to hours, failures down 82%, running costs down 97%) and the simulate-optimize-enforce loop, including the idea that small models can act as always-on reviewers in production.
The NVIDIA and Nebius comparison is the sharpest claim, and the vaguest in public. Plurai says post-trained models “for a global enterprise customer outperformed OpenAI and Anthropic frontier models on accuracy, at 40x and 44x lower cost, respectively.” The posts do not name the customer, the benchmark, the exact frontier models or the evaluation date. That should be read as a vendor result until more detail appears.
Pros and cons
For teams shipping agents
- Pro: A free starting point that plugs into tools many developers already use, aimed at the failures (bad tool calls, policy slips, leaks) that show up late in production.
- Con: A digital twin is only as good as the knowledge base, tools and policies you give it; missing a real edge case means the simulator will miss it too.
For buyers in regulated industries
- Pro: The company says it already worked with Fortune 500 and major financial design partners, which is the audience that needs this kind of rehearsal.
- Con: The headline accuracy and cost-multiple claims are not accompanied by a public, reproducible eval.
For the agent-tooling market
- Pro: More competition in pre-production testing is useful as companies move agents out of demos.
- Con: Simulation, evaluation and guardrail startups now crowd the same pitch, and buyers will have to test whether “free to start” becomes expensive once production feedback loops are on.
What’s still open
- Public details on the NVIDIA and Nebius accuracy comparison.
- How the free tier is limited, and what paid usage costs.
- Whether generated “fixes” are safe to apply without the same human review any other coding agent needs.
- How well the digital twin tracks a changing production environment.
About the analysis
Researched and written with AI assistance under the editorial responsibility of the i10X News team. The title image is an AI-generated image. Read our AI Use Policy. Spotted an error? Request a correction.
This independent i10x analysis is based on Plurai’s November 11, 2026 launch post, its press release, TechCrunch’s report and Team8’s investment note. i10x has not tested Plurai Simulation and has not verified the design-partner metrics or the NVIDIA and Nebius comparison.
i10x Perspective
The right instinct here is the car-industry one: do not let an agent meet a customer until it has failed in private, many times. The wrong instinct is to treat a launch-day cost multiple as settled science. If Plurai can make boring failures (a bad tool call, a leaked field, a policy miss) cheaper to find, it will earn a place next to eval harnesses already in use. The 40x claim can wait for a named benchmark.
