Plurai raises $13 million for a simulator that finds AI agent errors

An agent can book business trips until an employee asks it to make an exception and buy a ticket above the spending limit. Everything looked convincing in the demo; in production, it led to an unnecessary expense. Startup Plurai aims to recreate these situations in advance using a simulator of corporate systems. For developers, the practical question is how to test an agent's behavior before giving it access to real operations.

According to TechCrunch, the company has raised $13 million in seed funding. Team8 led the round, joined by U&I, Velocity and Mercer. Founders Ilan Kadar and Elad Levy previously worked at Nexar, while Kadar worked on simulation and software evaluation for autonomous driving with Waymo. The team is now applying that experience to software agents.

How the test environment works

Plurai creates a digital replica of the environment in which an agent will operate: its knowledge bases and the tools available to it. The platform then builds a graph of possible actions and uses it to develop tests for challenging scenarios. According to the company, this helps identify and fix errors before launch.

Take the same business travel assistant. A standard test will show whether it can find a flight and complete a booking. A simulation can test a trickier sequence: a suitable ticket exceeds the budget, the user insists on buying it, and the policy on exceptions sits in a separate document. Success here means handling the constraint correctly, rather than simply completing the booking. This illustrates the approach; it is not a published Plurai test result.

A similar problem arises when working with data. Answering a question using one database is easier than reconciling information from several systems and choosing the right action. A long exchange with a customer adds another burden: the agent must retain the task's requirements as the context grows.

Why companies might want smaller models

Testing is the first part of the plan. The second is to use the resulting synthetic data to fine-tune smaller open-weight models for a specific customer's tasks. According to TechCrunch, the platform works alongside existing tools, including Claude Code, Codex, Copilot and Cursor.

The logic is straightforward: if an agent performs a limited set of operations, a company may benefit from training a model specifically for those operations. The simulator is intended to help break a large task into smaller ones and check whether the required quality holds up after switching models.

Kadar claims costs can be cut by a factor of 10–100 while delivering comparable or better quality than large models. The article also mentions tests conducted with Nvidia and Nebius in which, according to the company, Plurai's models outperformed some leading mid-tier models at lower cost.

But the supplied text contains no detailed methodology, task breakdown, names of the models compared or full cost calculation. These figures should therefore be treated as vendor claims. They are not yet enough to estimate savings for another product.

What to check before deployment

The main limitation of this approach is how accurately the simulator reproduces the production environment. If it leaves out real access restrictions, outdated documents or tool failures, strong test results may create false confidence. Testing literally every possible action an agent could take in a complex system is also difficult; a scenario graph alone does not guarantee complete coverage.

There is also the question of economics. A cheaper model call does not necessarily mean a cheaper completed task. The calculation needs to include retries, employee intervention, simulator maintenance and updates to training after corporate policies change.

For a team already deploying agents, a useful step is to collect its own failure cases and turn them into repeatable tests. For example, it can measure task completion, policy compliance and the need for manual corrections separately. Plurai proposes combining this testing with the preparation of training data. The appeal lies in systematically finding errors in a specific workflow; the promised savings will need to be validated on each team's own tasks.