The calculus class connection
Phillip Li was 17 and Akira Tong was 14 when they met in a calculus class. It was the kind of early connection that would shape both of their careers: Li went on to become an executive at Amazon, Tong an engineer at Goldman Sachs and then Stripe. A decade later, the two reunited to start a company around a problem they'd both felt from the inside — AI agents that work in demos and break in production.
The problem was specific and expensive. Enterprises were rushing to deploy AI agents across Salesforce, Workday, email and the rest of their software stack, but the agents kept failing on the messy realities of corporate work: ambiguous records, overlapping workflows, permission systems, messages that may or may not have been sent. As Li told TechCrunch, an agent might need to decide whether two records created in different systems refer to the same company — the kind of judgment call that breaks automation in a live workflow.
The root cause, the founders argued, was how agents were trained. The industry standard was testing against stripped-down API mocks or stateless endpoints — environments that looked nothing like the real systems agents would face. It was like training a pilot in a video game and expecting them to fly a 747.
Digital twins, not API mocks
Arga Labs' answer is what the company calls digital twins: full-scale simulated replicas of enterprise software environments, cloned with permissions, webhooks and interaction logic intact. Where competitors settled for a stateless API endpoint, Arga rebuilds the entire program — effectively cloning Salesforce or Outlook so agents can train against realistic state without touching production data.
The technical claim is striking: Arga says it can clone any SaaS environment in under 12 hours with complete fidelity, and that its sandboxes can be reset between runs and seeded from a customer's actual stack. That reset capability is the key unlock for reinforcement learning, which needs to run the same scenario tens of thousands of times — something that's nearly impossible against live enterprise systems with rate limits and real customer data.
Early traction suggests the approach resonates. Customers completed more than 100,000 tests on Arga's platform over a 16-week period, a volume that reflects a genuine emerging requirement: as AI systems move from answering questions to performing multi-step actions across enterprise applications, testing has to get dramatically more realistic.
What makes the approach defensible is that fidelity compounds. Each enterprise system Arga learns to clone — each quirk of Salesforce permissions, each edge case in Outlook threading — becomes part of a library that makes the next clone faster and more accurate. Competitors starting from API mocks would need to rebuild years of accumulated realism. In infrastructure, the moat is often just depth: the unglamorous accumulation of edge cases that no demo ever shows.
The $10M seed
In August 2026, Arga Labs announced a $10 million seed round led by General Catalyst, with participation from Box Group, Emergence, Gradient and SV Angel. The company had previously raised a $500,000 round from Y Combinator, Comma Capital and Scribble Ventures. For a team of roughly four to five people, the $10M figure looks generous per capita — but it's consistent with the current AI infrastructure funding climate, where small elite teams command premium valuations.
The investor logic is straightforward. Every enterprise deploying AI agents needs training infrastructure, and the current tooling is inadequate. If Arga becomes the standard environment where agents learn to handle enterprise complexity, it sits at a chokepoint in one of the largest software transitions in decades. General Catalyst's bet is that the company that owns agent training owns a layer of the stack.
The Y Combinator connection mattered too. Arga went through YC's program, which gave the young company credibility with enterprise buyers who might otherwise hesitate to trust critical testing infrastructure to a five-person startup. In infrastructure, trust is the product — and YC's stamp helped manufacture it early.
Why it matters
Arga Labs illustrates a pattern in AI infrastructure: the winners build for the failure mode, not the demo. Every agent company can show a video of their system working. Far fewer can show it working ten thousand times in a row against a realistic Salesforce instance with conflicting data and weird permissions. Arga bet that the gap between demo and deployment was the business — and early customer usage suggests they were right.
The founding story also carries a lesson about co-founder selection. Li and Tong didn't meet at a hackathon or through a founder-matching program. They met as teenagers, built separate elite careers, and reunited with a decade of mutual trust already banked. When you're building infrastructure that enterprises bet their operations on, that kind of deep partnership shows up in the product's coherence.
Finally, there's the lean-team arithmetic. Five people, $10M raised, 100,000 tests in 16 weeks. Arga is a reminder that in AI infrastructure, leverage comes from the quality of the technical insight, not the size of the team. The digital twin idea is simple to state and brutally hard to execute — which is exactly the kind of idea that rewards a tiny team of exceptional engineers over a large team of good ones.
The playbook
- Replicate reality, not the API — Everyone else tested agents against stateless API mocks. Arga cloned the entire enterprise stack — permissions, webhooks, interaction logic — because that's where agents actually fail.
- Co-founders with history win — Li and Tong met as teenagers in a calculus class. A decade of shared context meant zero co-founder friction and total trust when the stakes got high.
- Let pedigree open doors — Amazon and Stripe on the founding resumes opened doors with enterprise buyers and investors. Pedigree doesn't build the product, but it buys the first meetings.
- Ship the metric before the revenue — 100,000 tests in 16 weeks gave Arga a usage metric that proved demand before revenue existed. For infrastructure startups, usage is the leading indicator.
- Stay lean at high valuation — Staying at 4-5 people while raising $10M kept burn low and optionality high. In AI infrastructure, a small elite team beats a large average one.