General Catalyst led Arga's seed round for Phillip Li and Akira Tong's four-person operation, which builds stateful replicas of business software.
By RuntimeWire Staff · Published
Primary source: TechCrunch
Why it matters #
Enterprise agents need repeatable places to practice actions, not another static benchmark. Arga is betting the resettable simulation environment becomes a key control point between capable models and production software.
Arga co-founders Phillip Li and Akira Tong have raised $10 million to build resettable practice environments for enterprise AI agents before those agents are allowed near live customer records, emails or payments.
TechCrunch reported on August 26th that General Catalyst led the seed round, with BoxGroup, Emergence, Gradient and SV Angel participating. Arga announced the financing less than a year after Li said he and Tong had started the San Francisco developer-tools maker.
Li and Tong met during their first-year calculus course at the University of British Columbia, when Tong was studying computer science and business and Li was studying neuroscience with plans to pursue human-biology research. Li later built an internal Amazon developer tool that he says saved over 10 recurring weeks of engineering work each year. Tong skipped high school, graduated from UBC at 19 and worked on fraud detection at Stripe after a stint as a quantitative analyst at Goldman Sachs. Tong also played Identity V professionally before leaving esports to finish his degree, according to Arga's Y Combinator profile.
Their route to agent-testing infrastructure began with an attempt to build an agent themselves. In a LinkedIn post published roughly nine months before the seed announcement, Li described Arga's original project as autonomous DevOps agents intended to handle on-call work. By Arga's Spring 2026 Y Combinator launch, Li and Tong had moved to the validation problem underneath that product: agents could generate and change software quickly, while teams still lacked safe places to test what those agents would do across real integrations.
That pivot put Arga on the infrastructure side of the agent boom. Li and Tong are selling the rehearsal room instead of another performer.
A reset button for business software
Arga builds stateful replicas of services including Salesforce, HubSpot, Slack, Stripe, GitHub, Gmail, Outlook, Notion and Google Workspace. Developers point an agent toward Arga's version of a service by changing a base URL, then seed the environment with mock or scoped production data.
The distinction matters because enterprise work rarely happens through a single clean API call. A sales agent may find one company represented differently in Salesforce and HubSpot, discover that another employee has already sent an email, and then decide which contact should receive the next message. Li framed the problem for TechCrunch with a basic identity question: "Can the agent correctly identify that these two are the same company?"
Arga's replicas preserve state between actions, including permissions, webhooks and side effects. Teams can then reset the environment and run the same workflow repeatedly for reinforcement learning, red-teaming, failure injection or evaluations. Arga's documentation describes repeatable scenarios, test runs tied to pull requests and evidence capture for the actions an agent took during a run.
That is a different technical problem from checking whether a model produced the expected text. An agent operating across business software can make a valid API request and still choose the wrong account, duplicate an action or expose data to the wrong user. Testing those failures requires a believable world around the model.
Arga's public product currently spans API-only replicas and services that reproduce both APIs and user interfaces. Arga says the system records provider calls, responses, latency, side effects and changes to service state. Its homepage displays engineering teams including Respan, Aemon, Bolto, Rho, Monaco, Weave and Slash as users of the product.
Arga has also put a visible business model around the infrastructure. Its pricing page lists a free tier, a $1,250-per-month Pro plan, a Team plan starting at $3,500 per month and custom enterprise pricing. Higher tiers add longer-running scenarios, automated test runs, custom twins and implementation support.
Investors are funding the agent practice layer
The seed syndicate gives Li and Tong backers across enterprise software and early-stage AI. General Catalyst led the financing. BoxGroup, Emergence, Gradient and SV Angel joined the round.
The timing reflects a rush to build training, simulation and evaluation infrastructure around agents. In June, Patronus AI announced a $50 million Series B alongside what it calls Digital World Models for complex digital workflows. Centific introduced an environments-as-a-service product for reinforcement learning in March. Veris AI says its own simulation product covers over 30 SaaS services.
Arga's wedge is the operational detail inside familiar business applications. Li and Tong are betting that agent developers will pay for replicas that behave closely enough like production systems to expose mistakes before deployment.
Fidelity will decide whether that bet works. A twin that misses a permission rule, webhook sequence or API quirk can produce a passing evaluation that means little in production. Each additional replicated service also creates another external product Arga must track as interfaces and behavior change. The breadth shown on Arga's homepage helps sales only if the individual replicas remain accurate.
Li and Tong have built their pitch around failures they encountered before Arga. Li saw the cost of validating changes to Amazon.com. Tong concluded at Stripe that staging environments lose much of their value when they drift from production. Their first plan to build autonomous DevOps agents then gave them the same problem from the customer's seat.
The $10 million round finances a direct response to that experience: build a controlled copy of the workplace, let agents make their mistakes there, and charge developers before those mistakes reach the real one.