The Agentic Learning Flywheel: How Holo3 Solved Computer Use
A new model just hit 78.85% on OSWorld-Verified, the leading desktop computer use benchmark. That's not the interesting part.
The interesting part is how it got there.
Holo3 didn't just train on more screenshots. H Company built what they call an "agentic learning flywheel"—a synthetic environment factory that generates enterprise workflows on demand, then trains agents to execute them.
This is the pattern that will define the next era of AI agents: not bigger models, but better training environments.
The Synthetic Environment Factory
Most computer use benchmarks are static. OSWorld presents a fixed set of tasks—open Excel, send an email, configure settings. Agents memorize the patterns.
H Company's approach is different. Their Synthetic Environment Factory builds entire applications from scratch using coding agents. You describe a scenario—"an e-commerce platform with inventory management and approval workflows"—and the factory generates:
- The application code
- Realistic data
- Multi-step tasks with verifiable outcomes
- Verification scripts that check the agent actually completed the work
The result: infinite training data that covers scenarios no human would bother to annotate.
The Flywheel Effect
The flywheel has three stages:
1. Synthetic Navigation Data – Using human and generated instructions, generate scenario-specific navigation examples. The key insight: you don't need humans to click through apps. You need them to describe what needs doing.
2. Out-of-Domain Augmentation – Programmatically extend scenarios. If the agent learned to navigate a shopping cart, can it handle a shopping cart with a broken payment flow? An unexpected popup? A login wall?
3. Curated Reinforcement Learning – Filter the generated data, then train with RL to maximize task completion. The synthetic environments provide clean reward signals—you either completed the task or you didn't.
The flywheel compounds. Each training run reveals edge cases. The factory generates new environments covering those cases. The model improves. Repeat.
Why This Matters for Production AI
The OSWorld benchmark is a proxy. The real test is whether agents can handle enterprise workflows—the messy, multi-application tasks that define actual work.
H Company built H Corporate Benchmarks to measure this: 486 multi-step tasks across E-commerce, Business software, Collaboration, and Multi-App setups.
The hard tasks require coordination across systems:
- Pull equipment prices from a PDF
- Cross-reference against employee budgets
- Send personalized approval or rejection emails to each person
This isn't navigation anymore. It's sustained reasoning across applications without losing state or intent.
The Cost Advantage
Holo3 achieves its scores with 10B active parameters out of 122B total. That's a fraction of GPT 5.4 or Opus 4.6.
The economics matter. An enterprise agent that costs $0.50 per task to run is very different from one that costs $5.00. The flywheel doesn't just improve capability—it improves efficiency.
The Next Frontier: Adaptive Agency
H Company calls the current milestone "interface mastery." The next phase is Adaptive Agency: models that learn to navigate entirely new, bespoke enterprise software in real-time.
This is the right framing. Today's agents work well on software they've seen before. The production reality is custom ERP systems, internal tools built on Airtable, workflows that change quarterly.
The synthetic environment factory points toward the solution. If you can generate training environments at scale, you can teach agents to generalize—first across known applications, then across applications that didn't exist during training.
The Pattern Behind the Pattern
There's a deeper lesson here. The companies winning at AI agents aren't just collecting more demonstrations. They're building infrastructure for synthetic data generation:
- Coding agents that write applications
- Verification scripts that prove completion
- Reinforcement learning loops that optimize for real outcomes
The model is the visible layer. The training environment factory is the moat.
Holo3 is a milestone. The pattern behind it—synthetic environments, curated RL, out-of-domain augmentation—will define who wins the agent race.
Holo3 weights are available under Apache 2.0 license. The free inference tier makes experimentation cheap. The question isn't whether this approach works—it's who will apply it to their domain first.