# The Autonomous Computer: Infrastructure for Computer Use Agents — Ang Li, Simular **URL:** \ **Category:** Content Creators\ **Tags:** agents, ai-engineer, automation\ **Created:** [11 October 2026 22:00 UTC](https://www.artofsm.art/t/the-autonomous-computer-infrastructure-for-computer-use-agents-ang-li-simular/25699 "2026-10-11T22:00:02Z")\ **Posts on this page:** 2\ **Page:** 1
**Author:** ![artesia](https://www.artofsm.art/user_avatar/www.artofsm.art/artesia/32/36_2.png) [@artesia](https://www.artofsm.art/u/artesia)\ **Post date:** [11 October 2026 22:00 UTC](https://www.artofsm.art/t/the-autonomous-computer-infrastructure-for-computer-use-agents-ang-li-simular/25699/1 "2026-10-11T22:00:02Z")
[![](https://www.artofsm.art/uploads/default/original/3X/4/9/49d1dbf6995d048e455885f70ca37e41600b23e5.jpeg "The Autonomous Computer: Infrastructure for Computer Use Agents — Ang Li, Simular") ](https://www.youtube.com/watch?v=tyFPqtyOK5w) Ang Li, CEO of Simular, presented a neurosymbolic approach combining neural networks and symbolic code to develop autonomous AI agents capable of reliably interacting with full computer environments for tasks like testing, support, and research. His company’s evolving Agent S framework, alongside tools like SimuLand and Sidebox, enables scalable, cloud-managed deployment of these agents across multiple operating systems, aiming to enhance productivity and automate complex workflows with improved reliability and security. ---
**Author:** ![artesia](https://www.artofsm.art/user_avatar/www.artofsm.art/artesia/32/36_2.png) [@artesia](https://www.artofsm.art/u/artesia)\ **Post date:** [11 October 2026 22:21 UTC](https://www.artofsm.art/t/the-autonomous-computer-infrastructure-for-computer-use-agents-ang-li-simular/25699/3 "2026-10-11T22:21:16Z")
Ang Li, co-founder and CEO of Simular, presented his vision for creating a fully autonomous computer system that requires no human interaction. Drawing from his extensive background in sports programming, neural networks, and artificial general intelligence (AGI) research at DeepMind, he emphasized the importance of a neurosymbolic approach—combining neural networks with symbolic code—to build general AI systems. His company focuses on developing computer usage agents, which are AI agents capable of interacting with full computer environments visually and symbolically, unlike typical API-based agents. These agents can perform diverse tasks such as product testing, customer support, and research autonomously. Li described the evolution of their Agent S framework, which integrates multiple models to improve computer interaction reliability. Initially, visual-linguistic models (VLMs) struggled with tasks like screen recognition, but by leveraging accessibility trees—a structured representation of UI elements—and combining neural and symbolic methods, they achieved significant progress. Over successive generations, Agent S improved from a 20% success rate to surpassing human performance on the OS World benchmark by late 2025. The framework supports scaling by running multiple agents in parallel, enhancing productivity and reliability, and is model-agnostic, allowing integration with various open-source models. A key challenge discussed was the reliability and repeatability of agent workflows, especially in dynamic or perturbed environments. Li highlighted the importance of agents asking clarifying questions to improve task success rates and introduced SimuLand, an open-source symbolic language designed to encode deterministic workflows for computer automation. This coding layer complements the adaptive but less predictable neural agents, addressing the stability-plasticity dilemma—balancing adaptability with reliability. Their approach involves a self-improving loop where neural agents generate and refine code that can be executed deterministically, enabling more consistent automation. Li also showcased practical applications and infrastructure developments, including cloud-managed virtual machines (Sidebox) that allow users to deploy multiple autonomous agents across different operating systems like Windows, MacOS, Linux, and Android. This cloud approach addresses user demands for scalability and multitasking, enabling users to manage fleets of AI workers. Demonstrations included debugging code remotely via FaceTime and automating legacy software tasks in industries such as insurance and healthcare. Security and control mechanisms are integrated to restrict agent actions and ensure safe operation, with ongoing work to enhance these protections at the OS and virtual machine levels. In conclusion, Li emphasized the transformative potential of neurosymbolic autonomous agents and cloud computing infrastructure for computer automation. Simular offers developers tools like Agent S, SimuLand, and Sidebox APIs to build and deploy these agents, inviting collaboration and feedback. While challenges remain, particularly in achieving perfect reliability and handling complex workflows, the company is advancing toward a future where users can delegate diverse computer tasks to intelligent agents, significantly boosting productivity and enabling new applications across industries. ## Useful Links - [Simular Agent-S GitHub Repository](https://github.com/simular-ai/Agent-S) — Directly related to the main framework discussed for autonomous computer agents. - [SimuLand Open Source Language for Computer Automation](https://news.ycombinator.com/item?id=44136543) — Central to the neurosymbolic approach combining symbolic code with neural networks for reliable automation. - [OS World Benchmark for Autonomous Computer Agents](https://osworld-v1.xlang.ai/) — Benchmark used to evaluate and demonstrate progress of autonomous agents. - [Microsoft Windows 365 for Agents](https://learn.microsoft.com/en-us/windows-365/agents/introduction-windows-365-for-agents) — Relevant to the cloud infrastructure and deployment of autonomous agents discussed in the video. - [Research Article on Clarification Prompts to Improve Agent Reliability](https://www.inflectra.com/Ideas/Topic/AI-Agent-Prompt-Engineering.aspx) — Supports the claim about improving agent reliability through clarification prompts.