Arga Labs just closed a $10 million seed round led by General Catalyst, with participation from Box Group, Emergence, Gradient, and SV Angel. The startup builds full-scale digital replicas of enterprise software like Salesforce, Workday, and email clients — not just stateless API endpoints, but complete clones with permission systems and web hooks intact. That distinction matters more than it sounds.
The core problem Arga tackles is one anyone who’s tried to deploy agents in a real enterprise has hit: you can’t reset Salesforce. Reinforcement learning demands running scenarios tens of thousands of times, but business software wasn’t built for that kind of iteration. CEO Phillip Li sketches a concrete example — a prospect created in Salesforce while a colleague reaches out separately through HubSpot. Can the agent recognize it’s the same company? Check whether an email was already sent? Identify the right contact? Those judgment calls are exactly where agentic systems still stumble.
Li’s framing is useful: this closes the reinforcement gap between coding and everything else. Coding tools advanced rapidly because developers already had infrastructure for deploying, reversing, and analyzing code — which made RL environments straightforward to set up. Business software never had that. Arga’s crash-test-dummy approach — replicating structure without the production system — could change the trajectory.
General Catalyst managing director Yuri Sagalov put it bluntly: ‘Having a repeatable sandbox environment is very important, and much more important with agents than it was with humans.’ He’s right. Human employees learn once and adapt. Agents need thousands of repetitions to internalize the same judgment. If Arga and its peers can make enterprise software as trainable as code, the bottleneck isn’t the models anymore — it’s the environment. And that’s a solvable problem.
