SimuLangのご紹介：デスクトップ全体のためのPlaywright

Simulangは、ブラウザ、ネイティブアプリ、OSレベルのワークフローを自動化するためのスクリプト言語で、AIエージェントによって記述されるように設計されています。Simulangをオープンソース化しました。今すぐ単一のコマンドでインストールできます。

Three years ago, my co-founder started Simular with one goal in mind: to build accessible general intelligence to liberate human labor, starting from the digital world. As agents take over mind-numbing computer work, humans are able to do what humans are best at and enjoy most: out in nature, home with family, making the creative leaps that unleash our full potential.

This week, Simular turns three. We are now a team of 30 across Palo Alto and Singapore, shipping weekly updates for our product, Sai. Meanwhile, our research arm keeps pushing the limits of computer-use agents. 

While the rest of the industry bet on chatbots or waited for software vendors to expose an API, we made a fundamentally different bet. We believe true general intelligence requires an agent that can interact with the digital world the way humans do – by operating a real computer, seeing the screen, clicking, typing, and navigating any GUI application. We set out as a research-first lab – publishing, open-sourcing, benchmarking honestly, and letting our results argue for us.

Looking back at the last 36 months, the trajectory of how this team has pushed us toward that horizon is staggering. We took two years of research and became the first in the industry to achieve human-level performance on OSWorld, a benchmark for computer use.

- Oct 2024 – Agent S (The first open source CUA): We put Simular on the map with a pure neural approach, introducing Experiential Learning with Retrieval, the Agent-Computer Interface (ACI), and Hierarchical Planning. Achieving a 20.58% on OSWorld, we set a new benchmark for GUI agents, notably outperforming Anthropic’s initial Claude Computer Use released a week later.

- Mar 2025 – Agent S2 (Harness beats a single model): We proved that architecture is its own axis of progress toward AGI. By pairing generalist neural planners with specialist grounding models, S2 surged to 34.5% on OSWorld (surpassing OpenAI and Anthropic again) and reached 50% on AndroidWorld, crossing from desktop to mobile.

- Dec 2025 – Agent S3 (Human-level computer use): Reframing variance as the core bottleneck to utility, we introduced Behavior Best-of-N (bBoN). S3 scaled performance to 72.6%—surpassing the 72.36% human baseline. Read more here.

We turned these research accomplishments into products that shape the lives of everyday users. This March, we launched our autonomous computer product, Sai. More recently, the advance in our neurosymbolic research has brought down token usage to 1/100 of industry average.

Crossing the human benchmark was a landmark, but relying purely on brute-force neural inference creates a clear ceiling: pure neural agents are too expensive, non-deterministic, and slow to run at global scale. We have developed our strongest conviction on neurosymbolic intelligence because it is the only architecture that mirrors the fundamental law of human labor. 

In the real world, work is naturally recurring. Human learning obeys the power law of practice: the more a task is performed, the faster, cheaper, and more automatic it becomes. Pure neural models violate this law – they re-reason through the exact same pixels and tokens every single time, burning compute continuously.

A neurosymbolic intelligence solves this completely through an Exploration ➔ Exploitation loop: (1) Neural Exploration: The agent uses multimodal reasoning to explore, navigate, and solve complex, novel digital environments. (2) Symbolic Exploitation: Once a task is mastered, it is compiled into deterministic code.

This is the bridge between frontier research and truly accessible intelligence. It gives us the reasoning capability of a human brain with the speed, low cost, and perfection of software.

Looking forward, we remain committed to our dual mission: grow Sai into an everyday tool for the mass market; we already see strong interest in areas like business development, healthcare and venture capital. In the meantime, our research and tech team keeps strengthening the infrastructure underneath it: the agent’s VM environment, scripting language, and 'brain.' 

All the while, Sai learns from people’s real needs, making autonomous computers better, more reliable, and cheaper to run. We are closer to our dream: a world where humans are freed from manual digital labor and empowered to define their own future.

Ang
CEO, Simular