Plurai Raises $13M to Simulation-Train AI Agents on Digital Twins

Plurai, a startup founded by two engineers who previously worked on autonomous driving, says it has raised $13 million in seed funding to build a simulation layer for AI agents, borrowing the playbook that helped self-driving cars reach public roads. The round was led by Team8, with participation from U&I, Velocity and Mercer, according to TechCrunch.

The pitch rests on a familiar complaint: agents are easy to demo and hard to trust. Benchmarks cited by TechCrunch show the gap. On LiveBench 2.0, Claude Fable 5.2 and ChatGPT-6 Astra succeeded on 82.9 and 75.6 out of 100 tasks respectively, and an evaluation by AI infrastructure company EON, run in a simulated corporate environment, found the strongest model tested, Claude Opus 4.8, completed tasks at 76.8% accuracy. Those numbers describe systems that work most of the time and fail often enough that someone has to watch them.

Why is agent reliability the problem Plurai is attacking?

According to TechCrunch, Plurai's co-founders Ilan Kadar and Elad Levi came from Nexar, which develops autonomous driving technology, and Kadar worked with Waymo to simulate and evaluate its end-to-end deep learning driving software. Kadar told TechCrunch that the pattern looked identical the moment agents arrived: "it is very easy to build a demo, but [hard] to get something that will be reliable and cost effective."

The failures TechCrunch describes are mundane rather than exotic. An agent booking travel can ignore a spending policy when a customer pushes back. An agent that answers reliably from one database starts to break when it has to combine several. A long customer service conversation piles up context until the agent takes a wrong turn. These are distribution problems: the system was never exercised against the awkward version of the task.

How does a digital twin of a company's systems get built?

Plurai creates a digital twin of a company's electronic systems, including the knowledge base and the tools an agent will have access to in production, according to TechCrunch. It then maps all the possible flows through those systems and uses that graph to generate evaluations and fixes for edge cases before the agent meets a real user.

That is the direct transplant from driving. Simulation mileage is how autonomy teams discovered the rare situations that a handful of road tests would never surface, and synthetic data is how they trained for them. Applied to agents, the equivalent of a rare road event is the policy exception, the cross-database query, the twenty-message thread.

The platform is positioned to sit alongside existing agentic coding tools rather than replace them. TechCrunch reports it works with Claude Code, Codex, Copilot and Cursor.

Why is post-training small models the bigger claim?

Kadar told TechCrunch that the simulation data matters more than the testing layer, because it can be used to post-train small, open-weight models tuned to one customer. The described loop is to take a large model, decompose a task into smaller pieces using the simulation, train an open-weight model on that, and then use the same simulation to verify the result is comparable to a frontier model at a fraction of the cost. Kadar's stated figure is one to two orders of magnitude of cost reduction.

TechCrunch reports that tests run with Nvidia and Nebius showed Plurai's models outperforming mid-tier frontier models at lower cost. Those are company-run results rather than independent evaluations, and no accuracy figures for the tuned models were published.

Reah Miyara, a Google DeepMind researcher who backed Plurai as an angel investor, told TechCrunch that the company's advantage is understanding production failure modes and generating synthetic data to tune against them. Her framing of the buyer's problem was blunt: IT leaders are realising that "frontier-capable models are out of distribution with my specific data, and the cost of these things isn't sustainable in the long term."

What does this mean for teams building agents now?

Three practical takeaways survive the funding announcement. First, evaluation is a design input, not a final check: if you cannot enumerate the flows your agent can take, you cannot know which ones are broken. Second, the failure modes worth testing are the ones your own data creates, not the ones a public benchmark measures. Third, the cost case for smaller tuned models depends on being able to prove quality did not drop, which is exactly the verification step Plurai is selling.

Kadar's own summary of the moat is that model capability is not the constraint: "no matter how smart OpenAI and Anthropic model will be — and they will be much smarter — they will not know your business and your policy," he told TechCrunch. TechCrunch also notes the physical-world parallel the founders are leaning on: general models are making robots more capable but not yet reliable enough for wide deployment, while task-specific models are already in the field gathering real data.

How could you explore this with Grafy?

Most teams cannot buy a simulation layer this week, but the underlying discipline is testable. If you are building a support or internal assistant, Bots is the place to define what it is allowed to answer from, and you can restrict it to your own content rather than open-ended generation. The related question of cost is handled differently in Grafy Chat, where you can put the same prompt to several models side by side and see where a cheaper one is good enough — the manual version of the decompose-then-verify loop Plurai describes.

For the research side, Reader can work through agent evaluation papers and benchmark reports you have collected, and the related desk coverage of agent memory verification and agent coordination studies covers adjacent ground.

Frequently asked questions

Did Plurai raise $13 million in seed funding?

Yes. According to TechCrunch, Plurai announced on Wednesday that it had raised $13 million in seed funding. Team8 led the round, with participation from U&I, Velocity and Mercer. The money is intended to build the company's simulation layer for AI agents, which generates evaluations and fixes for edge cases before an agent reaches production.

Does Plurai use simulation and synthetic data to prepare AI agents?

Yes. Plurai builds a digital twin of a company's electronic systems, including its knowledge base and the tools the agent will use in production, then maps the possible flows through them and uses that graph to generate evaluations and fixes for edge cases, according to TechCrunch. The same synthetic data is also used to post-train smaller open-weight models for individual customers.

Did tests with Nvidia and Nebius show Plurai's models beating frontier models?

TechCrunch reports that tests conducted in cooperation with Nvidia and Nebius showed Plurai's models outperforming mid-tier frontier models at lower cost. These are company-run results rather than independent benchmarks, and the report does not give accuracy figures for the tuned models or name which mid-tier models were used as the comparison.

---

Source: TechCrunch — AI — Plurai wants to help AI agents work better using the same approach that helped get self-driving cars to roads