Arga Labs raises $10 million for full-stack enterprise software simulation training environments

Problem: cannot initialize Salesforce

AI agents are expected to operate enterprise software, but they stumble when they must reconcile data across systems. Philip Lee, CEO and co-founder, cites a scenario where a potential customer appears in Salesforce while a colleague contacts the same firm through the AbsPath. The agent must recognize the duplicate, ensure a single email is sent, and decide which opportunity to use. Current agent platforms still struggle with this type of ambiguity.

Solution: digital twin with permissions and webhooks

Arga builds a digital twin of applications such as Salesforce, Workday and KlientiMail. The twin is more than a stateless API endpoint; it replicates the full permission model and active webhooks. Because the environment is under the company’s control, it can be reset, altered and run in many parallel instances. This enables reinforcement-learning at a scale comparable to code-level training without disrupting production systems.

Gap versus code-centric tools

Rapid progress in code-focused AI tools stems from mature infrastructure for deployment, rollback and code analysis, which supports complex RL environments for software. Enterprise applications lack such infrastructure. Arga aims to close that gap; if successful, the leap seen in code tooling could repeat in other business-software domains.

What investors say

Yuri Sagalov, managing partner at General Catalyst and head of the fund’s seed program, said, “The primary economic value of agents will come from use in business applications.” He added, “A sandbox that reproduces itself with precision is far more critical for agents than for humans.” General Catalyst led the round, with participation from Box Group, Amgen, Gradient and SV Angel.

What remains unknown

Arga has not released benchmarks showing measurable agent performance gains after training in its environments. It is also unclear how quickly a digital twin can be built for a new enterprise application or what maintenance is required when the source software updates. The company now has funding, a plausible thesis and a long list of potential customers, but real-world validation is still pending.