隆重推出 SimuLang：桌面版 Playwright
作者 Ang Li • 加利福尼亚州帕洛阿尔托 • 2026 年 4 月 23 日
Simulang 是一种用于自动化浏览器、原生应用和操作系统级工作流的脚本语言，旨在由AI代理编写。我们刚刚开源了 Simulang。您现在只需一条命令即可安装它：
.png)
That absence creates problems. AI agents, as the industry currently builds them, lack efficiency.
To be clear, this isn’t about AI’s capability; frontier labs have been breaking new grounds virtually every month. The problem is agent cost – measured in dollars, time, tokens, compute, and human oversight – and the fact that none of those costs gravitate toward the most efficient result. Just as how things in nature tend towards efficiency, computer autonomy should, too.
我们为何构建它
最近，我注意到我在电脑上的时间已大大减少到每天大约两个小时。一年前，我轻松到了八岁。不同之处在于，随着该行业最终开发出可以像人类一样观察和行动的代理，计算机使用代理（CUA）正在变得越来越好。
速度。每次操作耗时不到 50 毫秒——这只是查询本地 API 并执行点击所需的时间。无需图像捕获、上传或模型推理。一个 20 步的工作流可在不到一秒内完成。基于截图的代理在相同工作流中，每次操作需要 3 到 5 秒，这使得它们在规模化应用时慢 60 到 100 倍。
成本。Simulang 脚本在重放时不会消耗任何 token。您只需在首次编写脚本时（或 Sai 从自然语言生成脚本时）支付 LLM 推理费用。此后，每次后续执行都是免费的——没有 API 调用，没有云处理，也没有按次运行的费用。对于每天运行数百个自动化工作流的团队来说，这决定了方案是可行还是成本过高。这些并非渐进式改进。它们是选择正确抽象（语义元素而非像素、本地执行而非云推理、确定性引用而非概率性猜测）所带来的结构性优势。
Lastly, today's agents don’t practice because they never learn on the task. Agents of today are predominantly LLM-based. While an LLM model is great at solving novel problems, it’s wasteful on routine tasks. It doesn’t learn on the task but re-inferences from scratch every time. The LLM keeps doing the work without ever forming muscle memory, so it never gets more efficient at a task it has already done a hundred times. This turns into billable amnesia.
Lastly, today's agents don’t practice because they never learn on the task. Agents of today are predominantly LLM-based. While an LLM model is great at solving novel problems, it’s wasteful on routine tasks. It doesn’t learn on the task but re-inferences from scratch every time. The LLM keeps doing the work without ever forming muscle memory, so it never gets more efficient at a task it has already done a hundred times. This turns into billable amnesia.
What practice looks like in a machine
So how do we build agents that actually follow the power law of practice? The good news is we already have a working example of extreme efficiency sitting inside our brain.
The human brain runs on about 20 watts [3]. That is the power of a dim light bulb, and with it, a person can hold a conversation, recognize a face across a noisy room, plan a week of meals, and second-guess a decision from 10 years ago. No machine we have built comes close to that efficiency, namely the ratio of output to energy consumption.
It’s worth sitting with that before talking about machine intelligence. The brain did not become efficient by being powerful. It gets there by being cheap, and it gets cheap through practice.
An AI agent should work the same way. It shouldn’t pay twice for the same task, and it shouldn’t starve the model to save on tokens. Like muscle, learned and repeated actions can be handled by a small model and stored in small memory units. Not everything needs to route through a central LLM – the "brain" – every time. Intelligence doesn’t require brute force; small, structured, symbolic units can execute repeated work in code efficiently. That is the power law of practice: the first run is an expensive discovery, and every run after gets cheaper.
.png)
This is also the neurosymbolic approach: neural LLMs where novelty is needed for discovery, and symbolic, muscle-memory-like code for repeatable work. The neural half handles first encounters; the symbolic half is where practice happens. Together, they give agents efficiency overtime.
Simulang 的功能
你可能会问：
Simulang 脚本可以：
- 打开任何应用程序——浏览器、原生桌面应用、系统对话框、文件管理器。
- 读取可访问性树——将每个按钮、文本字段、菜单项和标签公开为结构化、可引用寻址的元素。
- 确定性交互——通过元素引用而非像素坐标进行点击、输入、选择、切换、滚动、展开/折叠等操作。
- 回退到视觉识别——当应用程序不公开可访问性数据时，Simulang 使用像素级视觉定位来查找屏幕上的元素。
