Blog

Prompt engineering 教会我们「好好说话」。Harness engineering（驾驭工程）教会我们别再指望说话。

Prompt engineering taught us to ask nicely. Harness engineering teaches us to stop asking.

每一个 AI agent 平台对同一个问题都有不同的答案：用户如何与你交互？ 有的给你一条命令和一个聊天循环。有的给你 80 个子命令。有的给你 TUI，有的给你 REPL，有的只是往 stdout 打印文本。

Every AI agent platform has a different answer to the same question: how does a user interact with you? Some give you a single command and a chat loop. Some give you 80 subcommands. Some give you a TUI, some a REPL, some just print text to stdout.

我在 AllClaws 项目里跟踪 30 个 AI agent 平台已经六个月了。我读过它们的源代码，比较过它们的架构，编目过它们的失败模式。但”读懂 agent 如何工作”和”看着 agent 工作”是两回事。于是我做了个不寻常的尝试：我组建了一个虚拟研究团队，让三个不同的 AI agent 平台在同样的研究任务上协作，然后跟踪它们产出的一切。

I’ve spent six months tracking 30 AI agent platforms for the AllClaws project. I’ve read their source code, compared their architectures, cataloged their failure modes. But reading about how agents work is different from watching them work. So I tried something unusual: I built a virtual research team where three different AI agent platforms collaborated on the same research tasks, and I tracked everything they produced.

2026 年 7 月，是 AI agent 平台走向成熟的一个月。不是能力上的成熟——它们早就足够强大了——而是在那些不那么光鲜、却真正区分 demo 与生产系统的工程工作上成熟了。上下文压缩（context compaction）成为平台的一等公民。安全加固从”事后想到”变成了发布阻塞项。中国生态展示了它作为拥有独立引力的平行宇宙。而支撑这一切跟踪的基准测试基础设施，平台覆盖率从 26% 跃升至 76%。

July 2026 was the month AI agent platforms grew up. Not in capability — they’ve been capable for a while — but in the unglamorous engineering that separates demos from production systems. Context compaction became a first-class feature. Security hardening moved from afterthought to release blocker. The Chinese ecosystem revealed itself as a parallel universe with its own gravity. And the benchmark infrastructure that tracks all of it leaped from 26% to 76% platform coverage.

每一个 AI agent 平台都附带一个完美运行的 demo。agent 读取一个 GitHub issue，写出修复，提交 PR。Slack 机器人总结一个频道，起草回复，然后等待批准。这些 demo 是真的，但它们也是谎言。

Every AI agent platform ships with a demo that works flawlessly. The agent reads a GitHub issue, writes a fix, opens a PR. The Slack bot summarizes a channel, drafts a reply, and waits for approval. These demos are true, and they are lies.

Western observers of the AI agent ecosystem see a world dominated by LangChain, CrewAI, AutoGen, and a handful of coding agents. This view misses an entire parallel universe. China’s AI agent ecosystem — shaped by the Great Firewall, government regulation, and a distinct set of domestic model providers — has produced at least 15 significant open-source projects with combined GitHub stars exceeding 350,000.

The Governance Problem

AllClaws Now Tracks 30 Platforms

Executive Summary

Executive Summary

OpenCode vs Claude Code

一句话

The Fork Is Real

The Problem: Your Agent Works, But You’re Still the Bottleneck

Executive Summary

Introducing awesome-agent-runtimes

Executive Summary

Executive Summary

引言

引言

执行摘要

引言

Executive Summary

We’re excited to announce a new capability for the allclaws project: automated tracking of AI agent platform updates. This system monitors 8 major AI agent platforms, identifies significant changes, and generates research reports automatically.

Welcome to the official blog for allclaws! This is where we’ll share updates, tutorials, and insights about our multi-agent AI framework.