AI said:
- Excellent documentation (SKILL.md is one of the best AI-agent skill docs I've seen — clear decision trees, organized by task, proper warnings)
- `htmlsnapshot inspect` auto-discovery is genuinely impressive for unfamiliar pages
Damn,这可能是2026年最被低估的开发者生产力工具💻🤖!
The harness is the bottleneck,喵个咪,同样的Claude,换个壳,内存省20倍,启动快245倍。
所以信息差不仅省钱,还省 token 啊,你如果觉得用Claude慢,其实不是Claude的问题,是这个harness的问题。
Claude Code单开一个会话吃掉230MB内存,启动一次要3.4秒,你说它慢,它也没法辩解,但换掉harness,同样的Claude,启动快245倍,10个并行会话总内存只用260MB。
这是一个叫jcode的开源项目做到的,作者1jehuang,Rust写的,GitHub repo就在那摆着,benchmark数据完整可查,没有任何营销成分。
有意思的是,它还支持一个叫Self-Dev的模式,代理可以直接改jcode自己的源码,编译,重载,用来迭代自己,这不是demo,作者每天都在把这套跑在生产上。
说白了就是,Anthropic把最好的模型卖给你,但没人说harness必须用他们的,开源社区的进化速度,从来不输大厂,只是他们不打广告。
当然风险是真实的,Anthropic明确禁止OAuth第三方接入,封号灰色地带,建议先拿API Key跑,不要用主号赌,但这个方向依然值得盯着。
Browser4 comes with a built-in personal assistant that is currently verifying the following points:
1/ The scripting engine and GitHub CLI already installed on your computer form the optimal Lobster System.
2/ The combination of scripts and file descriptors constitutes the most effective tooling framework. Systems like SKILLs are merely transitional products.
3/ From user stories to software feature decomposition, and then to implementation—this is an achievable path for AI. This is what we call the AI Software Factory.
4/ I have always advocated for migrating existing social organizations into the world of agents. Elements like agent protocols and skill descriptions are merely surface-level manifestations. Observe how communication happens in the real world—how contracts are signed, how window services are handled, and how responsibilities are defined for employees in companies and institutions—and the principles of agent collaboration will become clear.