Arga Labs has raised a $10 million seed round to build resettable digital twins of enterprise software, giving AI agents a controlled place to practice work before touching live systems. General Catalyst led the investment, with Box Group, Emergence, Gradient and SV Angel participating.
The startup is targeting a problem that appears after a model can generate plausible actions: enterprise work spans multiple applications, permissions and changing records, making realistic training and evaluation difficult to repeat.
Digital twins preserve the complexity of work software
Arga creates full-scale recreations of applications such as Salesforce, Workday and email clients. Instead of exposing only a stateless API endpoint, its environments reproduce permission systems, webhooks and application state. The goal is to let an agent encounter the interactions it would face inside a company.
CEO and co-founder Phillip Li gives an example involving two sales systems. One employee creates a lead in Salesforce while another contacts the same company through HubSpot. A capable agent must identify that the records refer to one organization, avoid sending a duplicate email and select the right recipient.
Such ambiguity is difficult to represent with a short list of test inputs. It emerges from the history and relationships inside several programs. Arga’s thesis is that cloning the environment produces a better training signal than reducing each application to isolated calls.
The source does not name customers, publish agent performance gains or disclose how closely every twin matches updates to the original software. The examples explain the intended use rather than proving production effectiveness.
Resettable sandboxes make reinforcement possible
Reinforcement learning improves an agent by repeating a task many times and rewarding successful strategies. That loop is straightforward when an environment can return to a known starting state. A live enterprise application cannot safely be reset after every attempt, especially when one test changes records, sends messages or triggers another workflow.
Arga compares its digital recreations to crash-test dummies: controlled substitutes that preserve the structures relevant to the exercise. Because the startup controls the environment, it can reset or modify conditions and run many copies simultaneously.
The system can also connect several twins, creating a simulated workday where tasks overlap across programs and knowledge sources. That breadth is important for agents expected to complete outcomes rather than operate one interface. It also increases the risk that a training environment omits an edge case present in the live organization.
Enterprise agents borrow a lesson from coding
Li argues that coding agents improved quickly because software development already has tools for deploying, reversing and analyzing code. Repositories, tests and reproducible build environments make it possible to evaluate thousands of attempts without damaging a production system.
Most business applications lack an equivalent training layer. Arga wants to close that gap so reinforcement techniques can be applied to sales, human-resources and communication workflows. General Catalyst managing director Yuri Sagalov says repeatable sandboxes are especially important for agents because much of their potential economic value comes from operating business software. That is an investor’s assessment, not an independent market forecast.
The seed financing reflects demand for infrastructure around agents rather than another general-purpose model. Before companies delegate consequential work, they need evidence about permission handling, duplicate actions, state changes and recovery from mistakes.
Arga’s controlled twins could make those failures easier to reproduce and correct. The open questions concern fidelity and maintenance: a sandbox is useful only if it mirrors the rules and interactions that matter in the live system. The startup now has capital to test whether its environments can remain accurate enough to turn enterprise-agent training into a repeatable engineering process.
The closer those replicas track production, the more credible their evaluations can become.
