Arga Labs Built the Rehearsal Room for AI Agents
Most software teams test on mocks that lie or real accounts they can't reset. Arga Labs spins up a copy of production every time you open a pull request - and lets agents fail there, not in front of your customers.
There is a moment every engineer knows and no one enjoys. The code passed every test. The pull request was clean. Then it shipped, touched a real payment or a real inbox, and something nobody simulated came loose. The usual explanation is short: the place we tested in was not the place we shipped to. Arga Labs, a San Francisco company from Y Combinator's P26 batch, decided that gap was the whole business.
Arga builds real-world sandboxes for software and the AI agents that increasingly drive it. In plain terms: every time a team opens a pull request, Arga spins up a temporary copy of their application that behaves like production. Only the services that changed get redeployed. Everything else quietly routes to the real thing. It sounds like a small plumbing decision. It is closer to the entire trick.
The founders, Phillip Li and Akira Tong, are betting that the next scarce resource in AI is not a smarter model. It is a safe place for that model to be wrong. An agent that can book travel, move money, or update a customer record is only useful because of the software it operates - and that is precisely where it tends to break. Arga wants to be the room where it breaks in private first.
01 / THE PROBLEMThe staging environment nobody actually has
Ask a room of developers whether they have a staging environment and most will say yes. Ask whether it truly mirrors production and the honesty rate drops. Staging tends to drift. Integrations get mocked, and mocks are guesses about how Stripe or Slack behaves - guesses that are fine until the real service returns something the guess never imagined. The alternative, testing against real accounts, brings its own headaches: rate limits, states you cannot reset between runs, and no way to test at the scale a busy system demands.
This was not a theory the founders read about. Tong interned at Stripe and came away struck by how much developer speed hinged on high-fidelity staging, and how few companies actually had it. Li, during an internship at Amazon, built an internal tool to automate messy engineering workflows that saved more than ten weeks of engineer hours a year across teams. Both had seen the same thing from different sides: the testing environment is where productivity is quietly won or lost.
The arrival of AI coding agents sharpened the problem into something urgent. A strong engineer paired with an agent can produce changes far faster than before. But if that agent cannot automatically test what it wrote against a realistic environment, the human becomes the bottleneck again - reviewing, running, and babysitting every change by hand. You can be a hundred-times engineer and still be, as Arga likes to put it, a one-times tester.
There is a second failure mode that shows up as teams grow. When several people are changing overlapping parts of a system at once, a shared staging environment turns into a traffic jam. One person's half-finished branch pollutes another's test run, and merging concurrent changes without breaking something becomes its own full-time job. Scoping a fresh environment to a single pull request sidesteps that: each change gets its own clean world, and no one has to negotiate for the staging server.
02 / THE PRODUCTTwins, and the art of a convincing fake
Arga's core idea is the twin. A twin is a stand-in for an external service - Stripe, Slack, GitHub, Gmail, Salesforce - that speaks the same language as the real one. It exposes the same API endpoints, works with the same SDKs, and fires the same webhook events. The difference is that a twin is isolated, seedable, and resettable. You can fill it with a scenario written in plain English, run your workflow against it, and wipe the slate clean for the next test. No rate limits. No real customer on the other end.
Around the twins sits the sandbox. When a pull request lands, Arga deploys only the services that changed. Dependencies like databases and Redis run as in-memory sidecars, so a test can read from the production database while writing everything to the sidecar - nothing corrupts real data. If a team prefers total isolation, they can route all of it to the sidecar instead. The design gives teams fidelity where it matters and safety everywhere else.
Then comes the part built for the age of coding agents. A developer, or an agent acting on their behalf, can prompt Arga through the web app, an API, a command-line tool, or MCP to generate tests for a pull request, run them, and stream back the results and logs. The agent sees what was tested, what broke, and which specific function or call caused the failure. It can take that, go fix the code, and try again - so the pull request is already passing by the time a person looks at it. Arga has shown this running with Claude Code reading its logs directly.
Live twins span Discord, Dropbox, GitHub, Google Calendar, Google Drive, HubSpot, Salesforce, LinkedIn, Linear, Notion, Slack, Stripe, Box, Jira, and Gmail, with Attio, Outlook, and Twilio on the roadmap. Relative bars are illustrative, not survey data.
03 / THE DIFFERENCENot a mock, not an eval
Arga is careful about what it is not. It is not a mocking library, where you hand-write a fake response and hope it stays true. And it is not an evaluation platform that scores an agent's answers. It sits somewhere more practical: the infrastructure that lets an agent take a real action in a fake but faithful world, then watch what happens. The distinction matters because taking actions - charging a card, sending a message, moving a file - is where agents create both their value and their risk.
| Approach | Fidelity to prod | Resettable | Runs at scale |
|---|---|---|---|
| Mocked integrations | Low - hand-guessed | Yes | Yes |
| Testing on real accounts | High | No - state sticks | No - rate limits |
| Shared staging | Drifts over time | Hard to isolate | Merge conflicts |
| Arga twins + sandbox | High - same API/SDK | Yes - between runs | Yes - parallel |
The competitive field is a mix of old and new. There are service-virtualization and API-mocking tools that developers have used for years. There are preview-environment and sandbox platforms. And there is the fresh crop of agent-evaluation startups. Arga's claim is that none of them, alone, gives an agent a resettable, high-fidelity place to actually act. The bet is that as agents move from writing text to taking actions, that missing piece becomes the one everyone needs.
04 / THE BUSINESSMoney in, seven weeks later
The model is usage-based software. A free tier lets developers try pre-built twins and short sandbox runs. A Team plan starts around $1,000 a month, scaling with usage and team size, and adds long-running scenario agents, CI test workflows, saved scenarios, and shared access. An Enterprise tier layers on custom twins, unlimited validation, on-premises hosting, single sign-on, and SOC 2 compliance for larger buyers with stricter requirements.
Early demand has been notable for an infrastructure company. Arga reported roughly $40,000 in monthly recurring revenue within seven weeks of launch, from customers paying with their own budgets rather than pilot credits. In tooling, that kind of number tends to say less about marketing and more about how sharp the underlying pain was. The company launched publicly in April 2026 and was later named among the standout startups from Y Combinator's Demo Day in a VC roundup.
05 / THE FOUNDERSA 14-year-old, a fencer, and a shared problem
The origin story is unusually specific. Li and Tong met in first-year calculus at the University of British Columbia. Tong was 14 at the time; Li had come in studying neuroscience with plans to research human biology before pivoting to computer science within a year. Tong skipped high school and finished his degree at 19, worked as a quant at Goldman Sachs in Hong Kong, and then wrote fraud-detection software at Stripe. Li, a former fencer on the Canadian Junior National Team, went on to build developer tools at Amazon before a stint as a founding engineer at an AI startup.
What holds the two threads together is that both founders arrived at the same conviction from different jobs: that the quality of a company's test environment quietly sets the ceiling on how fast its engineers can move. Arga is the attempt to raise that ceiling for everyone, and to hand the same lever to the agents now writing a growing share of the code.
06 / THE MARKETReliability follows every wave
Arga frames its own position with a tidy observation: every technology wave builds a reliability market next to it. The web got monitoring and error tracking. The cloud got observability and infrastructure-as-code. Mobile got crash reporting and device farms. Agents, the argument goes, will get a testing and validation layer - a place to prove they behave before they are trusted with real systems. Arga wants to be the trust layer that every agent passes through on its way to production.
Whether that layer becomes a large standalone category or gets folded into the platforms agents already run on is the open question. For now, Arga has a working product, a growing library of twins, paying customers, and a clear thesis about where software reliability is heading. The company is small and early. But the problem it picked is old, specific, and widely felt - which is usually a better foundation than a clever product looking for a use.
The moat, if there is one, is probably the twin library itself. Each new integration Arga rebuilds faithfully - matching not just the happy-path responses but the odd errors, the webhook timing, the pagination quirks - is a corner of the real world an agent can now rehearse in. That work is unglamorous and slow, which is exactly what makes it hard to copy quickly. A roadmap that adds Attio, Outlook, and Twilio next hints at a company planning to widen that surface one careful service at a time rather than chasing breadth for its own sake.
For teams shipping agents into anything that matters, the pitch is concrete. Point Arga at a pull request, let it build a faithful copy of your world, and find out what your code or your agent does when it meets something it did not expect - while the only thing at stake is a sandbox you can reset. That is the quiet promise underneath the twins: fewer surprises where surprises are expensive.