Plurai opens AI agent testing with simulated orders and payments

Plurai opened its AI agent testing tool, Plurai Simulation, to all developers on November 11. The product sets up simulated versions of services such as order databases and payment systems, then runs an agent through tasks that would normally use them. The startup had offered the platform to selected customers and partners. It announced a $13 million seed round led by Team8 on the same day.

Development teams can use the tool to see where an agent mishandles a task before it reaches customers. Plurai released plugins for Claude Code, Codex CLI and GitHub Copilot CLI. Its documentation calls for an initial browser sign-in, and the company lists Cursor support as forthcoming. In the default setup, the simulation intercepts actions that would write to live systems and carries them out on test records.

A refund changes the test order

Plurai illustrates the approach with a customer reporting a duplicate charge. The order lookup and refund tools work on the same set of test records, so a lookup after a refund shows the change. If the agent refunds both charges when it should refund only one, the final order record exposes the mistake even if the customer gets a plausible reply. The example comes from the company’s technical explanation.

To build a test, the software identifies the tools an agent calls and the information those tools read or change. It generates a customer with a goal, starting records that match the situation and checks for the expected outcome. Teams can supply a product specification or an operating policy to narrow the test. Plurai’s guide says tests of an agent’s ability to retrieve answers from its knowledge base depend on adding the relevant material.

The report puts the conversation and the agent’s actions side by side. At a failed turn, it shows the tool the agent called, what came back and whether the response came from the simulation or a connected system. By default, the final run plays each scenario three times and flags cases that produce mixed results separately from those that consistently pass or fail.

Replaying the point where an agent failed

Teams that already have an agent in service can use a production record to restore a conversation and its state just before a failure. If an agent misapplies a policy partway through a call, the test starts with the history that led to the decision. Teams can also generate new cases from the agent’s code before they have production records and keep both types of cases in one suite.

Plurai’s report distinguishes possible bugs in the agent from problems in the test setup or mock data. It flags issues for further investigation when the evidence is insufficient. Developers choose which proposed code changes to apply, then can run the same cases again. A confirmed failure can become a regression test that runs against later versions.

The agent’s code runs on the developer’s machine, while Plurai hosts the simulated services. Its data documentation says the conversation, tool calls and tool results go to Plurai’s servers during the test so the service can play the customer and score what happened. Exact replay uses the original production conversation and is unavailable in anonymized mode. The company separately documents a workflow for using anonymized production records to inform other tests.

The company says preparing and checking a first suite can take hours. Its guide estimates 10 to 60 minutes to set up the simulated services the first time. Later builds can reuse the environment. A change to the agent’s tools or data sources calls for a new scan, while testing a proposed fix can use the existing suite.

As of November 13, Plurai’s pricing page listed two agent setups, seven scenarios, three full runs and 10 hours of sandbox time on the free plan. One full run includes multiple attempts per scenario. The free allowance pauses once it is exhausted. The Base plan costs $60 per month and includes 20 full runs and 30 sandbox hours per month.