Arga Labs is building a better way to train enterprise AI agents

Making AI agents behave inside real business software has turned out to be harder than plenty of companies hoped. Arga Labs is betting that the fix starts before deployment, with training environments that let agents practice on software that looks and acts like the real thing.

The startup announced a $10 million seed round on Wednesday, led by General Catalyst with Box Group, Emergence, Gradient and SV Angel also in the mix. Its target is enterprise software: Salesforce, Workday, email clients and the like. Instead of stopping at a stateless API endpoint, Arga builds what amounts to a digital twin of the whole program, including permission systems and web hooks.

That matters because enterprise work is full of ambiguity. Philip Li, Arga’s CEO and co-founder, points to a scenario where one person creates a lead in Salesforce while a colleague separately reaches out through HubSpot. An agent has to figure out whether those are the same company, whether an email has already gone out, and who should get the follow-up. That kind of judgment is exactly where current agentic systems still stumble.

The usual way to improve a system like that would be reinforcement learning, hammering the same task over and over until the good behavior wins out. But enterprise software is awkward to reset and even harder to clone. Arga’s pitch is that a recreated environment solves that problem: it can be reset, modified and run in many copies at once, which makes it possible to train agents across multiple systems instead of one isolated app.

There’s a reason coding tools have pulled ahead. Software development already has mature ways to deploy, reverse and inspect code, so RL environments are easier to build there. Most business software doesn’t have that infrastructure yet. Arga is trying to create it, and General Catalyst’s Yuri Sagalov says that repeatable sandbox environments are becoming more important as agents do more of the work inside business applications.