Arga Labs Raises $10M to Train Enterprise AI Agents
Fundraise Insider tracks newly funded startups each week and delivers verified sales leads of C-suite contacts straight to your inbox, so you reach the right people while the funding is still fresh. See pricing.
Arga Labs has announced a $10 million seed round led by General Catalyst, with Box Group, Emergence, Gradient and SV Angel joining.
The company was founded in 2025 and works from San Francisco. That puts it among the dense concentration of agent tooling startups now operating in California. Chief executive Phillip Li previously held an executive role at Amazon, and chief technology officer Akira Tong came from Stripe and Goldman Sachs.
Getting agents to work in practice has proven harder than many companies assumed, and a small group of startups has formed around testing and training them before they touch production. Arga builds training environments for enterprise software including Salesforce, Workday and email clients. Where most testing setups settle for a stateless API endpoint, Arga clones an entire program, permission systems and web hooks included, producing a digital twin rather than a stub.
Li illustrates the problem with a scenario most sales teams would recognize. A prospect creates a lead in Salesforce while a colleague reaches out separately through HubSpot. Can the agent work out that both records point to the same company, confirm the email was sent only once, and pick the right recipient between two opportunities? Agentic systems handle that kind of ambiguity poorly.
Ordinarily an agent would learn a task like that through reinforcement learning, running the scenario tens of thousands of times and keeping only the strategies that succeed. Enterprise software makes that nearly impossible. There is no straightforward way to reset Salesforce or Outlook between runs, let alone clone them.
Arga’s answer is a re-creation of the software that mirrors its structure the way a crash test dummy mirrors a person. Because the company controls the environment completely, resetting or altering it is trivial, and many instances can run at once so agents learn the interactions between programs rather than one application in isolation. The goal is to reproduce an entire work environment, with tasks that span multiple tools and knowledge systems.
One way to read this is as an attempt to close the reinforcement gap between coding and everything else. AI coding tools improved quickly in part because mature tooling already existed for deploying, reversing and analyzing code, which made reinforcement learning environments easy to build. No equivalent exists for most business software. Once it does, systems should get considerably better at operating those programs.
The team remains small, around four to five people, and it is aiming at mid to large engineering organizations either building their own coding agents or rolling out third party autonomous workflows.
Companies selling into newly funded AI infrastructure teams can source verified contacts through B2B software sales leads gathered from recent rounds.