This side of the river is the old world. The other side is AI. I want to ferry us all across.
RSS FeedFull-stack developer for 24+ years. I formerly led a 400-person engineering team, and I'm still a core maintainer of several large open-source projects. Everything here is something I've actually built, run, or broken: AI engineering practice, open-source dynamics, infrastructure, and security.
中文读者请进 简体中文版 ——每篇文章都有中英双版。
Featured
I tore down HumanLayer's source code: it leashes your coding agent, but sessions move to its cloud
HumanLayer open-sourced its approval layer for coding agents. I read the source: the engineering is real, but there is a simpler path using the CLIs' own built-in doors.
Recent Posts
Meta's new Muse Ember costs $0.14 per task, while Haiku 5.5 costs $0.41
Meta released Muse Ember on November 9. It shares Haiku 5.5's per-token price list ($0.10 input and $0.50 output per million tokens), yet Artificial Analysis measured $0.14 per finished task for Ember against $0.41 for Haiku 5.5, because Ember writes about a third as many output tokens per task and its cached input is about 60% cheaper. The catch: a single request over 200,000 input tokens bills the whole order at 3 times the price. Judge a model by what one finished task costs, not by the price list.
Fable 5.2 scored 49.1% at coding, but Anthropic itself held back its GPU kernel-writing skill
Anthropic released Claude Fable 5.2 and Mythos 5.2 on November 9, two names for the same model. Fable 5.2 scored 49.1% on FrontierCode 2.0 and costs 20% less, but classifiers inside it block high-risk cybersecurity and most research-biology requests and hand them to older models — and in training, Anthropic deliberately held accelerator kernel optimization and language model architecture to roughly Opus 5.5's level. The fallback meant to prevent abuse hands requests to an older model that injection breaks into more easily: 2.05% when Fable 5.2 answers itself, 5.17% when Opus 4.8 answers instead.
Meta open-sourced the design system behind 13,000 apps: a component library friendlier to agents
Meta open-sourced Astryx, the design system that ran inside Meta for eight years across more than 13,000 apps. Its API, docs, and CLI were designed together under one set of conventions, so people and AI agents build interfaces the same way — down to prop definitions documented field by field in the source, and a swizzle command that ejects a component's full source into your own project. It is still in beta, so the honest test is to hand an agent a real page and count how many rounds it needs to get it right.
DeepSeek Harness hit 1.54M daily users in five weeks, with 35,000 plugins already built
DeepSeek's agent app Harness went from a seven-day average of 271,000 daily users on September 30 to 1.54 million on November 4 — about 5.7 times in five weeks — and third-party plugins already number around 35,000, with about 60% of active users having built or installed at least one. The moat is moving from the model to the plugin ecosystem around it. The faster that ecosystem grows, the more concrete the security side gets: Harness's own safety notice says the software has not passed a security audit. Before installing a plugin, check which files it reads and which tools it calls.
Turn a Codebase into a Graph: Is floorplan Any Good?
A hands-on test of floorplan v0.6.3: four agent sessions took turns working on the pytest repository for a day, then a fresh session that had taken no part tried to reconcile the day from the records floorplan left behind alone. The map covers all 284 files without missing one, and locating files and single-session records are genuinely good; but the recent-work record is replaced wholesale on every update, so after a full day only 1 of the 4 tasks could be fully recovered. Enough for one agent checked session by session; not enough for several agents reconciled together at day's end.
Open WebUI 0.13.1 moves coding agents onto your own computer, but not without a paired desktop
A hands-on install test of Open WebUI 0.13.1, ruled pain by pain: coding agents now work on your own computer through the Tether desktop companion while approval stays in your chat. File extraction and admin-enforced temporary chats are solved; reconnect resume and the Auto approval mode are only partly solved; the service-tier choice is one agent with a single tier and its description dropped in the interface. The flagship agent path has not been run end to end, so existing users should upgrade, while new users should first check whether they are who that path was built for.
