Company profile · 4 min read
Halluminate: resettable computer-use environments for agent training
Halluminate provides managed sandboxes, datasets and evaluations for teams training and testing browser and computer-use agents.
Cofounder of Lazyweb (a16z Speedrun). Previously grew subscriptions and built native ads at Duolingo. Writes First1000 (90,000+ subscribers).
Published · Updated
Halluminate builds training data, evaluations and resettable environments for teams teaching AI to use browsers and business software. Its current public wedge is financial services, but the underlying buyer is any model lab or enterprise that needs computer-use behavior tested without letting an agent touch a live customer system.
What it does
The current Halluminate site describes reinforcement-learning environments for investment banking, private equity and consulting workflows. The YC profile describes a broader suite: managed sandboxes modeled after systems such as Salesforce, Slack and ticketing software, proprietary datasets and expert-annotated evaluation services. The launch explains why the sandbox matters: real sites are noisy, dynamic and hard to reset, with authentication, ads and other side effects mixed into the test.
The buyer is a frontier-model lab or enterprise AI team building browser or computer-use agents. Halluminate says paying customers include leading computer-use model labs and the two largest browser-agent companies in the space; that is company-reported customer context. Its offer is less “give us a benchmark score” and more “give us a reproducible place to find failure modes, train against them and compare the next version.”
Why I’d look closer
The advantage is experimental control. A sandbox can reset state, run in parallel and expose a clean reward or failure signal without sending an agent to a real CRM, ticket queue or financial system. Founder context fits the problem: Jerry Wu previously led product and research at Capital One Labs and co-authored patents; Wyatt Marshall is described as a data and software engineer across early startups.
The tradeoff is realism. A simulated Salesforce or financial workflow is only useful if its permissions, data dependencies, error messages and incentives resemble the customer’s real environment. A model can learn to win the sandbox while failing on the live system it was meant to operate.
What I’d ask
Which financial workflows and sandboxes are available today? Can a customer define its own roles, records, tool contracts and reset conditions? How are annotations audited? Does a benchmark measure recovery from partial failure and permission denial, or only task completion? Can the same trajectory be replayed after the environment changes?
My editorial take
Shortlist Halluminate if computer-use reliability is blocked by unsafe or irreproducible testing. Start with one owned workflow and compare sandbox failures against a controlled fixture, keeping real credentials and customer data out of the environment. The product is strongest when it makes the failure surface legible; its risk is optimizing for a clean simulation nobody actually uses.
Quick facts
| Field | Sourced detail |
|---|---|
| Product | Managed RL environments, datasets and expert evaluation for computer-use agents |
| Buyer | Model labs and enterprise AI teams, with current public focus on financial services |
| Environments named | Salesforce, Slack, ticketing and financial-services workflows |
| Pricing | Not published in the checked pages |
| Main question | Does the sandbox preserve the failure modes that matter in the live workflow? |
Sources checked
| Source | Checked |
|---|---|
| YC company profile | 2026-09-19 |
| Halluminate homepage | 2026-09-19 |
| Halluminate YC launch | 2026-09-19 |
Cohort context
Halluminate is listed in Summer 2025. In our 2026-09-18 directory snapshot, 112 of 166 listed companies in that cohort have YC’s primary industry label B2B (67.5%). This is a current-directory comparison, not an original intake count or a performance ranking. Nine-cohort dataset.
Public website snapshot
Observed 2026-09-19T16:18:06.434Z in raw homepage HTML. This records visible metadata and advertised links, not agent execution or product quality.
| Signal | Homepage observation |
|---|---|
| Product description metadata | Observed |
| Canonical link | Not observed in this response |
| H1 or H2 heading | Observed |
| Typed structured data | Observed |
| Docs/developer link | Not observed in this response |
| Pricing link | Not observed in this response |
| llms.txt link | Not observed in this response |
| Markdown alternate | Not observed in this response |
Public observations · Collection method. Missing links here do not establish that a capability or file is absent elsewhere.
About the author
Cofounder of Lazyweb (a16z Speedrun). Previously grew subscriptions and built native ads at Duolingo. Writes First1000 (90,000+ subscribers).
I cofound Lazyweb and publish Mudpie. This is an owner-written publication, not an independent testing organization. Research notes distinguish observations, sourced reporting and editorial judgment.
First1000 ↗ · X ↗