Building a simulated internet so AI agents can rehearse the real one - before they touch anything that matters.
A five-person shop in San Francisco teaching software to practice. The stage is fake on purpose; the failures are cheap, repeatable, and, for once, measurable.
There is a boring problem hiding underneath one of the most hyped ideas in technology. Everyone wants an AI that can use a computer - book the flight, fill the spreadsheet, close the ticket, navigate the enterprise software that nobody enjoys navigating. The models are getting good at this. But "good at a demo" and "reliable in production" are separated by a gap, and the gap is not really about intelligence. It is about practice. An agent that learns by clicking around the live internet is slow, flaky, and impossible to grade. Did it actually book the right flight, or did it just say it did? Halluminate's answer is to stop using the real internet as a classroom and build a synthetic one instead.
The company, part of Y Combinator's Summer 2025 batch, sells two unglamorous things to the labs and enterprises building agents: realistic, resettable sandboxes, and high-quality data to train and grade against. Its flagship effort is called Westworld - which, if you have seen the show, is a fairly on-the-nose name for a simulated world where you rehearse safely before the stakes get real. Inside Westworld are synthetic versions of the apps agents most need to master: things shaped like Salesforce, Slack, ticketing tools, storefronts, booking sites. The data is procedurally generated, so it stays stable across runs. The whole thing is deterministic and runs offline, with no dependency on the actual internet, which means an experiment you run today produces the same world tomorrow.
"Training AI agents to use computers, browsers, and software is one of the highest-potential opportunities for AI - but it's still unreliable."
Halluminate, Launch HNHere is the mechanism, because the mechanism is the whole company. An agent is handed a task and a verifier. The task might be: book a flight from San Francisco to New York with a specific set of filters. The verifier is a small program that checks the final state - does the booked flight, expressed as JSON, match what was asked? That yes-or-no answer becomes a reward signal, which is the fuel for reinforcement learning. The industry has a clunky acronym for this: RLVR, reinforcement learning with verifiable rewards. The clever part, and the part that is genuinely hard, is the word verifiable. Anyone can generate tasks. Reliably grading whether an agent completed one, at scale, deterministically, is the bottleneck. Halluminate is selling the pickaxe for that specific mine.
If you run a model lab or an enterprise trying to ship an agent, Halluminate gives you a place to train it, a way to grade it, and a diagnosis when it fails. Four pieces:
A fully simulated internet of synthetic consumer and enterprise apps. Agents practice economically valuable tasks - flight booking, financial modeling, data reorganization - and every attempt is verified programmatically to produce an RL reward.
Fully managed, parallelizable, resettable environments modeled after popular systems like Salesforce, Slack and ticketing software. Reproducible enough to trust the numbers you get back.
Curated, high-quality training datasets and evaluation benchmarks for computer- and browser-use agents - the ground truth that RLVR quietly depends on.
Human-generated data and analysis that tells you not just whether an agent failed, but where and why - the part demos never show you.
A synthetic app - a storefront, a booking site, a CRM - loads with procedurally generated but deterministic data. Same seed, same world, every time.
The agent receives an objective ("book SF to NYC with these filters") and starts clicking, typing and navigating like it would on the real thing.
A programmatic verifier inspects the final state as structured JSON and decides, unambiguously, whether the task was completed correctly.
That pass/fail signal becomes a reinforcement-learning reward. Repeat at scale, in parallel, offline - and the agent gets measurably better.
Halluminate partnered with the agent startup Yutori to build and test Westworld: 100 tasks across five simulated environments. An agent trained directly inside the simulators posted these numbers. Reproducible worlds make scoreboards like this mean something.
Source: Halluminate x Yutori Westworld benchmark. Figures as reported by the company.
Jerry Wu and Wyatt Marshall met their first week at Cornell studying computer science, and have lived and worked together for more than seven years since. That is longer than most startups exist. When people say "pick your co-founder carefully," this is the version where they actually did.
Previously led product and research at Capital One Labs, where he launched one of the first AI agents in banking and co-authored three patents. Studied CS and Economics at Cornell.
A Cornell Milstein Scholar who ran large-scale data engineering for two early-stage NYC startups before co-founding Halluminate.
"Westworld is a fully-simulated internet made up of synthetic versions of the most common consumer and enterprise apps."
HalluminateThere is an old, reliable strategy in technology booms: while everyone else races to build the flashy end-user thing, quietly sell the infrastructure everyone needs to build it. Halluminate is a shovel company. Its customers are the ones with the compute and the models - foundation-model labs and browser-agent companies - and its product is the thing they cannot easily buy off the shelf: faithful simulators and verifiable rewards. The company has said it works with leading computer-use labs and browser-agent firms, and it is open-sourcing environments through its Westworld GitHub repository, betting that giving away the gym is a good way to become the standard place people train.
The competition is real - academic web benchmarks, in-house simulators at the big labs, and the broader data-labeling establishment all overlap with pieces of what Halluminate does. The wager is that "faithful and deterministic" is genuinely hard to build, that labs would rather buy it than maintain it, and that the agent era needs honest scoreboards more than it needs another impressive video. On the funding side, the numbers reported publicly are modest and inconsistent - a seed round in the low hundreds of thousands by some records - which for a five-person YC company at this stage is roughly what you would expect and not the interesting number to watch. The interesting number is 86%, and whether the next one is higher.
Unlock real advances in browser and computer-use AI by giving builders realistic, resettable environments and quality data - so agents learn to do valuable work reliably, not just impressively.
B2B infrastructure with two revenue streams: RL simulators and sandboxes, plus human-generated data services - training datasets, evaluation, and error analysis.
Y Combinator (S25), Antigravity Capital, Orange Collective, Batch Ventures, Team Ignite Ventures and Transpose Platform Management among the reported investors.
Jerry Wu and Wyatt Marshall found Halluminate in San Francisco to tackle the training bottleneck for computer-use agents.
Public launch via Y Combinator and Launch HN: "Simulating the internet to train computer use." Reported seed round.
Westworld v1 published alongside a Yutori partnership benchmark - 86.0% average across 100 tasks, 100% on the simulated store Megamart.
Reporting compiled from public sources: Y Combinator, Launch HN, the Halluminate website and blog, and press. Funding figures are approximate and vary by source.