Browser and computer-use AI is the most-watched capability in the model lab roadmap — and the most fragile in deployment. When OSWorld launched in 2024, the best model completed 12.24% of routine desktop tasks against a human baseline of 72.36%.^[23] WebArena showed the same gap on multi-app web workflows.^[1] a16z's framing is identical: "current agent offerings more closely resemble advanced RPA tools than true autonomous systems."^[13] The gap has since closed at exactly the rate labs bought environments — which is the point.

The reason isn't model capacity. The reason is that labs train agents on whatever data they can scrape — public web recordings, OSS web benchmarks, synthetic prompts — and then fine-tune them on a few thousand hand-curated trajectories. None of that mirrors the production stack the agent will actually run against. The Salesforce instance, the Slack workspace, the ServiceNow ticket queue, the QuickBooks reconciliation flow — these are the surfaces the customer cares about, and they look nothing like the public web.

Live-web training makes it worse. Real websites can't be reset. They rate-limit aggressively, ban automation, and punish exploration. They change layout overnight. They cost real money when an agent buys the wrong flight. Every lab knows this; every lab has tried to build a stand-in internally; every lab has discovered that environment authoring is a product problem masquerading as a research problem.

Computer-use agents are now first-party products. OpenAI shipped ChatGPT Agent. Anthropic shipped Claude Computer Use. Google launched Project Mariner. Browserbase shipped Director.^{[3] [4] [13] [15]} The category went from research demo to flagship product in twelve months. Every one of those launches has the same next-step problem: making the agent reliable on the long tail of enterprise apps.

Post-training is where capability now comes from. The o-series, Claude 3.5 Computer Use, and ChatGPT Agent were all RL post-training stories. Agents can't learn knowledge work through real-world trial and error — they need custom environments that faithfully simulate reality and reward success.^[12] The output of post-training is no better than the environment it was trained against. The benchmark record makes the causality legible: OSWorld went from 14.9% (Claude 3.5 Sonnet, October 2024) to 38.1% (OpenAI's CUA, January 2025) to 61.4% (Sonnet 4.5, September 2025) to 84% (Opus 4.8, May 2026) — each jump an RL-against-environments release, none of them a bigger base model.^{[24] [25] [26] [28]}

The money is following the bottleneck. Hyperscaler AI capex hit $344B in 2025.^{[5] [6]} Meta paid $14.3B for Scale.^[7] Surge held funding talks at a $25B+ valuation.^{[9] [20]} Mercor quintupled to $10B in eight months on a $450M run rate.^[19] And the spend has gone explicitly environmental: The Information reported Anthropic leadership discussing more than $1B on RL environments over the next year, with typical lab contracts running six to seven figures per quarter.^{[16] [17]} Felicis calls RL environment platforms "foundational infrastructure for anyone looking to train generalist AI workers."^[14]

Design task → simulate → instrument → verify → train → review. Halluminate runs the same loop the best in-house lab teams run, packaged as infrastructure. Customers pick a workflow, Westworld stages a sandbox for it, checkers score every episode, and Athena's reviewers triage the failures into reward models. The loop already shows up in customer numbers: one customer reported a ~20% improvement in date-picking performance after training against Halluminate's flight-booking simulator.^[29]

Delivery is verification-gated. A lab runs a smaller model against the environment; a second model grades the runs and flags errors — did the agent invoke the right tool for the task; then a larger model stresses it again. Failures route back to Halluminate to debug before the contract is accepted and paid. Increasingly the training target is the decision layer — the reasoning that picks the right tool, not the execution of the tool itself — which makes the verifiers, not the simulator, the scarce artifact. An environment that survives acceptance is lab-grade by construction.

Reusable modules bend the curve. Authentication, billing, search, form fills, ticket queues, notification banners — the same primitives show up in every enterprise UI. Each new environment shares more of its scaffolding with the last. The catalog grows superlinearly to engineering hours, the same way Applied Intuition's scenario library did in AV.^[11]

Interoperable with the agent stack labs already use. Westworld plugs into popular agent frameworks, browser automation infra (including Browserbase),^[15] and the standard training pipelines labs run. No new framework to adopt — Halluminate becomes the env layer beneath whatever the customer already runs.

Why static data ran out of room. Pre-training was about scraping the world. Post-training is about practicing in it. A static label set can teach a model what a "good" outcome looks like once. An interactive environment teaches it how to recover when something unexpected happens — and recovery is most of what agent reliability turns out to be.

The pattern the labs already ran. The earliest lab investments in environment infrastructure — Procgen, DeepMind's StarCraft sandbox, the OpenAI Gym lineage — were research bets that environments are the scarcest resource in RL.^[8] The same pattern is repeating one level up the stack: instead of toy gym tasks, the scarce environments are real enterprise workflows.

Why environments and the data they produce are sold together. Every agent episode against a Halluminate environment produces a trajectory, a checker score, an annotator review, and a reward signal — exactly the training input the labs need for the next model release. Customers buy environments today and end up paying for the resulting trajectory corpus on every retrain.

Frontier model labs. Three to five labs control the high-end of the post-training spend, and the line item is now public: Anthropic leadership has discussed over $1B on RL environments in a year, and typical environment contracts run six to seven figures per quarter.^{[16] [17]} Each lab is staffed with a small team trying to produce computer-use environments fast enough to keep up with the agent roadmap. Halluminate is already in active pilot with one of them, targeted to convert this quarter. Sustained AI capex creates adjacent demand for training and eval infra to help labs realize their model investments.^{[5] [6]}

Serious agent companies. Browser Use, Yutori, Manus, Browserbase Director, and the next wave of agent products all need the same thing the labs need.^{[13] [15]} Their differentiator is reliability in the customer's actual stack — which means training and grading against environments that mirror that stack. Halluminate sells the same product on the same loop.

Enterprise. The medium-term buyer is the enterprise platform team deploying internal agents. Vertical functions — marketing, finance, sales, HR — all require company-specific tuning against company-specific surfaces.^[13] The same Westworld + Athena loop powers internal agent evaluation before the agent ever touches production. Halluminate has since planted its flag on the highest-value vertical first: the company now leads with RL environments for financial services — Excel modeling, investment banking, private equity, and consulting workflows — where task value per episode is highest and domain expertise is the barrier.^[30] The demand side has gone vertical too: Anthropic now ships finance-agent templates with Excel and Moody's integrations — financial institutions are roughly 40% of its top-50 customers — and Rogo's $160M Series D at $2B (April 2026, a 2.7× step-up in under four months) priced what agentic finance is worth.^{[34] [35]} Whoever trains those models needs finance-grade worlds to train them in.