How Arga Labs is Making AI Agents Ready for the Real World

Arga Labs, building real-world sandboxes for testing and training AI agents, led by Phillip Li and Akira Tong, has raised a $10 million seed round led by General Catalyst, joined by BoxGroup, Emergence, Gradient, and SV Angel.

As AI agents become more capable, traditional software testing is becoming insufficient.

Traditional software is largely deterministic: given an input, developers can define the expected output and test against it. Agents are different. Their non-deterministic behavior and ability to make decisions autonomously are central to their usefulness, but also make them harder to test.

Agents interact with real software by calling APIs, reading and writing data, responding to events, handling permissions, retrying failures, and navigating complex workflows. Once tools enter the loop, the testing surface extends beyond the model itself.

Benchmarks such as WebArena and WorkArena already evaluate agents in realistic, executable environments, while standards such as MCP are making more software directly accessible to agents. As agents gain access to more tools and systems of record, the consequences of failures also become greater.

The answer isn't simply more test cases. Agent testing requires realistic simulations of the environments in which agents will operate, giving them a safe version of the real world to navigate, fail in, and learn from before they are trusted with real systems.

Real-World Sandboxes.

Arga is built on the premise that AI agents require a fundamentally different approach to testing. Instead of shallow mocks or live production systems, Arga provides real-world sandboxes built from high-fidelity twins of the external software agents interact with.

Agents can interact with these environments through APIs, CLIs, and MCP interfaces, while the sandboxes reproduce the underlying service behavior, including authentication, authorization, permissions, mutable resources, webhooks, timing, failures, and retries.

Because agent actions can change software state, Arga also captures the internal state of each SaaS environment, giving teams visibility into how agents affect integrated systems while avoiding production risks, rate limits, and state accumulation.

Arga can now clone a SaaS product in under 12 hours with 100% fidelity in backend functionality and behavior. Customers have already run more than 100,000 twin instances in 16 weeks.

From Tests to Environments.

Most testing tools mock API endpoints and provide stateless environments. Arga instead replicates the internal state of services alongside the CLI and MCP interfaces preferred by agents. The twins also reproduce permissions, authentication, webhooks, asynchronous behavior, and tiered access.

Arga then connects these twins into deterministic scenarios, multi-app environments with shared, synchronized state. For example, an e-commerce support scenario can combine Slack, Stripe, and Jira twins with the same users and synchronized order data.

This unified environment is critical because agents rarely operate within a single application. They execute sequential and parallel workflows across multiple systems, creating failure modes that isolated SaaS testing cannot capture.

Arga measures this behavior through ArgaBench, a multi-app agent benchmark powered by its software twins. Early testing found that frontier models, including Anthropic’s Claude Fable 5 and OpenAI’s GPT 5.6 Sol, struggled with cross-application verification even when completing most of a task. In one example, both models failed to change a product price in Stripe and independently cross-check the metadata in Notion.

Reliable long-horizon agent behavior across multiple software systems remains an unsolved challenge. Arga is building the validation infrastructure needed to make general-purpose agents reliable beyond individual workflows.

Building for What Comes Next.

Today, agent testing focuses on systems that choose among a defined set of tools. As agents become more capable, they may discover public APIs, understand their schemas, authenticate, call them, combine multiple services, and operate across systems never anticipated by developers.

Arga is using the new funding to accelerate R&D on technology that can generate high-fidelity SaaS twins in minutes rather than hours. As agents gain the ability to discover and interact with arbitrary APIs, testing infrastructure will need to dynamically generate the environments around them.

The company is also developing an evaluation platform to support agent testing and training. Because Arga controls the state of the simulated services, it can capture the traces, anomalies, and failures generated during agent interactions, helping teams identify weaknesses and improve performance.

Agent failures should be discovered in simulation, not production.

𝐑𝐞𝐦𝐨𝐭𝐞 𝐇𝐢𝐫𝐞 𝐖𝐢𝐭𝐡 𝐔𝐬: 𝐌𝐞𝐧𝐥𝐨 𝐓𝐚𝐥𝐞𝐧𝐭 helps technology companies hire exceptional remote talent from India. Build your team with carefully curated engineers, product leaders, designers, GTM professionals, and more. Learn More At: https://www.menlotimes.com/menlo-talent