Building reliable AI agents is proving far more challenging than many businesses anticipated. Fortunately, assistance is emerging: a fresh wave of startups is developing improved methods to evaluate and refine agents before they go live, especially when it comes to handling the intricacies of modern corporate systems. Arga fits squarely into this category. On Wednesday, the company announced it had secured $10 million in seed funding, led by General Catalyst, with additional backing from Box Group, Emergence, Gradient, and SV Angel.

Arga specializes in constructing training environments for enterprise platforms such as Salesforce, Workday, and email services. While most testing setups rely on a basic stateless API endpoint, Arga builds a comprehensive digital twin of the software—essentially cloning the entire program, complete with its permission structures and web hooks. This approach enables far more effective training of agents across interconnected systems. The startup’s CEO and co-founder, Philip Li, demonstrated the product using a scenario where a potential client creates a lead in Salesforce while a teammate separately reaches out via Hubspot. “Can the agent correctly identify that these two are the same company?” Li asks. “Are they able to check whether or not they’ve only sent the email once? Are they able to identify who to send the email to out of the two opportunities?”

Agentic systems still stumble on this kind of ambiguity, and Li views Arga’s tools as essential for driving improvement. In theory, an AI agent could be trained for such tasks using reinforcement learning (RL)—running the scenario tens of thousands of times and keeping only the strategies that succeed. However, the reality of enterprise software makes that level of testing nearly unworkable. There’s no straightforward way to “reset” a system like Salesforce or Outlook when you need to replay the same situation, let alone duplicate it. Arga’s answer is to digitally recreate the software by mirroring its architecture, much like a crash test dummy stands in for a human body. Since Arga has full command over the environment, the recreation can be easily reset or tweaked. The company can also spin up multiple environments simultaneously to train agents on how different programs interact.

The broader vision is to simulate a person’s entire workspace, where tasks overlap across various platforms and knowledge bases. Think of it as closing the reinforcement learning gap between coding and other domains. One reason AI coding tools have progressed so rapidly is that we already possess sophisticated infrastructure for deploying, reversing, and analyzing new code. That makes it straightforward to set up RL environments for coding, enabling us to test and train AI systems on increasingly complex programming challenges. Those same tools don’t yet exist for most business software—but once they do, AI systems are poised to become dramatically more adept at using those programs, potentially transforming other industries just as they’ve reshaped software development.

Yuri Sagalov, managing director at General Catalyst and head of the firm’s seed investing program, sees a rising demand for agentic testing solutions like Arga. “Having a repeatable sandbox environment is very important, and much more important with agents than it was with humans,” he notes.