Arga is a startup that builds training environments for enterprise AI agents, announced a $10 million seed round led by General Catalyst with participation from Box Group, Emergence, Gradient, and SV Angel.

The company’s core product is a digital twin of enterprise software such as Salesforce, Workday, and email clients, allowing AI agents to be trained in a realistic but fully controlled setting.

Unlike typical testing environments that offer stateless API endpoints, Arga replicates the entire software—including permission systems and web hooks—so agents can encounter and learn from complex, real-world scenarios.

The CEO, Philip Li, gives an example: an agent might need to determine whether a lead created in Salesforce and a separate outreach in Hubspot refer to the same company, and then decide on the correct action without sending duplicate emails.

Such ambiguity is hard to handle, and training agents to manage it normally requires reinforcement learning with thousands of repetitions, which is impractical in live enterprise systems because you cannot easily reset or clone them.

Arga‘s approach makes the environment resettable, modifiable, and parallelizable, enabling large-scale training.

The idea is to close the ‘reinforcement gap’ that exists for business applications, similar to how coding tools already benefit from sandboxes and version control.

General Catalyst‘s Yuri Sagalov notes that the economic value of agents lies in using business applications, and having a repeatable sandbox is crucial—more so for agents than for human users.