Making AI agents work in practice is a lot harder than many companies expected — but there's help on the way. A new crop of startups is finding better ways to test and train those agents before they get deployed, particularly on the complexities of the modern enterprise. Arga Labs is one such company, which announced its $10 million seed round not long ago. The round was led by General Catalyst with participation from Box Group, Emergence, Gradient, and SV Angel.
But here's the thing most people miss: training enterprise AI agents isn't just about having enough compute or the right prompts. It's about the environment. Most testing environments settle for stateless API endpoints. Arga builds a full-scale digital twin of the program, effectively cloning an entire enterprise program with permission systems and web hooks intact.
Take the example CEO and co-founder Phillip Li gives. A prospective client creates a lead in Salesforce, while their colleague reaches out separately through 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 struggle with this kind of ambiguity. And Li sees Arga Labs' tools as critical to helping them improve.
Normally, the agent could be trained for a task like this through reinforcement learning: essentially, running the scenario tens of thousands of times and letting only the successful strategies through. But the nature of enterprise software makes that scale of testing nearly impossible. There's no easy way to "reset" a system like Salesforce or Outlook when you need to run the same scenario again, much less clone it.
Arga Labs' solution is to create a digital re-creation of that software — replicating its structure the way a crash-test dummy replicates a person. Because Arga has complete control over the environment, it's simple to reset or modify. The company can also run many environments at once, training agents on the complex interactions between different programs. The idea is to replicate a person's full work environment, with specific tasks overlapping between different programs and knowledge systems.
You can think of it as a way to close the reinforcement gap between coding and other applications. Part of the reason AI coding tools have advanced so quickly is that we already have sophisticated tools for deploying, reversing, and analyzing new code. Those tools make it much easier to set up RL environments for coding, which lets us test and train AI systems on increasingly complex coding tasks.
Those tools don't exist for most business software — yet. But once they do, you can expect AI systems to get much better at using those programs, revolutionizing other industries the same way they've revolutionized coding.
The Reinforcement Gap That Won't Go Away
The problem isn't hypothetical. When an agent misidentifies two leads as the same company, the consequences multiply fast — duplicate emails, confused customers, broken workflows. Most teams cope by cutting corners: narrower task scopes, synthetic data that doesn't match real usage, or simply crossing their fingers and deploying anyway. None of those scale. enterprise software doesn't lend itself to the kind of "reset and retry" loops that make reinforcement learning practical in other domains.
Code has it easy. Modern tooling can spin up fresh repositories, run test suites, and tear down environments in seconds. Try that with Salesforce, Outlook, or a custom ERP, and you're looking at change windows, data integrity risks, and compliance headaches. Arga Labs' Phillip Li put it plainly: the enterprise doesn't have a reset button. That's why the company built a digital twin — not a mock, not a proxy, but a fully operational replica where every permission, every hook, every data relationship lives at exact parity with the production system.
When Salesforce and HubSpot Collide
The Salesforce-HubSpot example Li describes cuts to the heart of what makes agentic AI tricky in the real world. Two separate systems, two separate records, both pointing to the same customer. An agent that can't tell them apart will send two marketing emails instead of one, or worse, overwrite data in ways that are hard to undo. Digital twins eliminate that ambiguity by giving the agent a single source of truth — a sandbox where the same business rules, the same data model, the same identity resolution logic all operate under Arga's controlled environment.
Phillip Li's question — "Can the agent correctly identify that these two are the same company?" — is really a question about whether the agent has learned the right identity resolution strategies. In a digital twin, the answer is tractable because the environment is repeatable. Run the scenario a thousand times, note which strategies succeed, and you have real training data, not guesswork.
From Seed to Sandbox: The $10M Question
Arga Labs didn't announce just a concept. The $10 million seed round, led by General Catalyst with participation from Box Group, Emergence, Gradient, and SV Angel, signals that the market sees something worth betting on. Yuri Sagalov, who runs General Catalyst's seed program, has been watching the space and agrees the need is growing. He told TechCrunch, "I think that a lot of the economic value from agents is from using business applications," and added, "Having a repeatable sandbox environment is very important, and much more important with agents than it was with humans."
The investment also validates the team's approach. Building digital twins at scale isn't cheap, and the participation of established venture firms suggests they see a defensible business in helping enterprises de-risk agent deployment. Box Group's interest makes particular sense — the company has long been bullish on tools that improve how software is built and operated, and Arga's approach fits that thesis.
The Coding Comparison That Isn't
It's worth dwelling on the coding analogy, because it's become the default shorthand for this kind of discussion. AI coding tools have advanced rapidly, in no small part because the industry already has sophisticated tooling for deploying, reversing, and analyzing new code. That tooling makes it easy to set up reinforcement-learning environments for coding tasks — you can branch, test, and merge without risking production systems.
Most business software doesn't have that luxury. You can't branch a Salesforce org the way you branch a Git repository. You can't reset an Outlook inbox with a single command. Arga Labs' digital twin effectively provides that reset button, but it does more than just clone the interface. It clones the logic, the permission boundaries, the web hook integrations — the whole stack that makes enterprise software what it is. Run a scenario a hundred times, modify a parameter, run it again — each iteration is faithful to the production environment, which means the training signals are real, not artifacts.
What General Catalyst Sees
Yuri Sagalov's insight that repeatable sandbox environments matter more with agents than with humans gets at a subtle but important point. Human workers can learn on the job, make mistakes, and correct course over time. AI agents often get one shot at deployment, and if they've internalized the wrong patterns, the cost of correction can be significant. A repeatable sandbox changes the economics: teams can iterate freely, test edge cases, and stress-test edge cases without worrying about production fallout.
That's the promise Arga Labs is betting on, and the seed round suggests the market is ready to find out whether the promise holds up at scale.
The Road Ahead
Arga Labs is not the only startup tackling the training problem. A new crop of companies are all exploring different angles — some focusing on simulation, others on synthetic data, still others on improved observability during agent execution. What sets Arga's digital twin approach apart is the fidelity: the company isn't approximating the enterprise environment, it's replicating it down to the permission systems and web hooks that make enterprise software behave the way it does.
Whether that level of fidelity is enough to move the needle on agent reliability remains to be seen. But the need is undeniable. As more enterprises experiment with AI agents, the ones that can train those agents safely and at scale will have a significant advantage. Arga Labs' digital twin is one answer to that problem, and the $10 million seed round proves there's capital willing to bet the approach works.
Teaching AI agents to navigate enterprise software without breaking things is hard. Really hard. But the startups finding better ways to test and train those agents before deployment are changing the math, and the rest of the industry is watching closely to see what sticks.
Source: TechCrunch, "Arga Labs is building a better way to train enterprise AI agents," Aug 26, 2026, by Russell Brandom. See also the YC Spring 2026 Demo Day coverage on Arga Labs here.
Further Reading
- YC Spring 2026 Demo Day: Defense Tech, AI Agents, and Infrastructure Dominate Record Valuations
- Target's Real Competitive Advantage: Why Agentic Data Platforms Outshine Raw AI Models
