This startup raised $10 million to give AI agents a practice arena before they touch real business software

Key points

• Arga, a San Francisco-based startup, raised $10 million in seed funding led by General Catalyst, announced Wednesday.
• The company builds digital twins, exact software replicas, of enterprise tools like Salesforce and Workday so AI agents can train on them safely.
• Current AI agents struggle with real business software partly because there is no easy way to reset those systems after each test run.
• Arga's approach borrows from the method that helped AI coding tools advance so quickly: a controllable environment that can be run thousands of times.

AI agents, software that can carry out multi-step tasks on its own such as booking meetings, updating customer records or sending emails, are being built by practically every major tech company right now. Getting them to work reliably inside a real business is turning out to be much harder than anyone advertised.

The core problem is training. To get good at a task, an AI agent needs to attempt it tens of thousands of times, fail, adjust, and try again. That process is called reinforcement learning. It works brilliantly for coding, where you can deploy a change and then instantly roll it back. It falls apart with business software.

Why is training AI agents on business software so hard?

You cannot easily undo a Salesforce entry or reset an Outlook inbox after every test run. Without a reset button, you cannot run the same scenario thousands of times, and without that repetition, the agent never properly learns.

Arga, founded by Philip Li, is trying to fix that. The company builds a digital twin, a full working replica of a piece of enterprise software, complete with its permission settings and automated triggers. Think of it as a crash-test dummy that stands in for the real program. Because Arga controls the copy entirely, the team can reset it instantly, modify it freely, and run many copies at once to test how an agent handles tasks that span several different tools simultaneously.

Li walked through a concrete example. A sales team using both Salesforce and HubSpot, two popular customer-relationship platforms, might receive enquiries from the same company through both systems at the same time. Can the agent recognise they are the same company? Does it know whether an email has already been sent? Does it pick the right person to contact? Today, most agents get tangled in exactly that kind of overlap.

Arga's environments are designed to throw those overlapping, ambiguous scenarios at an agent repeatedly until it learns to handle them.

Who funded this, and why do they care?

General Catalyst led the $10 million seed round, with Box Group, Emergence, Gradient and SV Angel also investing, first reported by TechCrunch AI. Yuri Sagalov, a managing director at General Catalyst who runs the firm's seed programme, said the need is real and growing.

"A lot of the economic value from agents is from using business applications," Sagalov said. "Having a repeatable sandbox environment is very important, and much more important with agents than it was with humans."

What does this mean for people who use these tools at work?

Nothing changes today. Arga sells to the companies building AI agents, not to end users directly. But the gap it is closing matters to anyone whose employer is deploying an AI agent to handle their customer data or internal workflows. Better-trained agents make fewer embarrassing, or costly, mistakes.

The analogy to coding is worth keeping in mind. AI coding assistants got good partly because developers already had strong tools for testing and reversing changes. Once similar infrastructure exists for business software, AI agents handling sales, HR and finance tasks are likely to improve at a similar pace.

Arga is betting that moment is close.

Common questions

What is a digital twin in this context?

A digital twin is a working software replica of a real programme. Arga's version copies tools like Salesforce in enough detail that an AI agent cannot tell the difference, while giving engineers full control to reset or alter it between training runs.

Does this affect my company's actual data?

No. The point of Arga's approach is that agents train on the replica, not the live system, so real customer records and emails are never touched during the training process.

When might better-trained AI agents reach ordinary workplaces?

Arga is at seed stage, meaning it is early. Wider deployment of reliably trained enterprise agents is realistically a matter of years, not months.