codex 里的 agent 之前全跑 Astra/high，用量烧得很快。GPT-6 的 Sol 和 Luna 出来以后， 我拿七种模型和 effort 组合在真实 repo 上跑了三轮：自己的 monorepo、工作上三个 repo 的跨项目题、 还有盲测复现历史 PR。代码能写对的比想象中多，但在 agent 自己的集成 runner 里把真实调用路径也跑通的少很多， 好几份自测全绿的提交其实没修好。文末是最后定下来的分工表和几个派活规则。

opus 5.5 出来当天，我拿自己 11 个 repo 里的 22 个真实提交做了一轮 medium 对 high 的配对测试： 两档完成度一样，high 慢三分之一、多花三成多 token，但盲评里多抓出几个测试查不出来的真 bug。 再对一下 Artificial Analysis 的 Intelligence Index 数据，最后把默认档定成了 high。 文末按不同用法给了推荐。

Every Codex agent I ran was on Astra/high, and usage burned fast. When GPT-6 Sol and Luna landed, I had seven model and effort combinations run through three rounds on real repos: my own monorepo, cross-project tasks in three work repos, and blind replays of historical PRs. More routes wrote correct code than I expected; far fewer also got the real caller path passing in their own integration runner, and several submissions with green self-tests weren't actually fixed. The final routing table and delegation rules are at the end.

On launch day I replayed 22 real commits from 11 of my own repos, running Opus 5.5 at medium and high effort side by side. Both finished every task; high took about a third longer and 35% more tokens, but blind judges credited it with catching several bugs the tests couldn't see. Cross-checked against the Artificial Analysis Intelligence Index, I made high my default. Recommendations by usage profile at the end.

Reading Anthropic's September misuse report through the economics of agent workflows, the limits of account bans, and the gap between observed activity and proven harm.

读 Anthropic 九月滥用报告，聊 agent 为什么能让小团队持续做复杂的坏事，以及模型拒绝、人工审批和账号封禁各自能管到哪里。

上一篇让几个 claude 账号自己换着用，人还是拴在电脑前。这一篇把 agent 全挪到一台常开的机器上，手机随时连上去看进度、派新活，笔记本合盖它也接着跑。三个开源件各管一段——herdr 管 session、hcom 管通讯、Moshi 管手机——外加账号自动轮换兜底，凑成我目前跑得最顺的移动端 vibe coding 方案。这篇讲怎么搭，也讲踩过的坑。

Last time I taught a few Claude accounts to rotate themselves — but I was still chained to my laptop. This time I cut the last cord: the agents live on an always-on box, and my phone connects in to watch and steer them. Close the lid, leave the house, lie in bed — they keep running. Three open-source pieces each own one layer — herdr the sessions, hcom the messaging, Moshi the phone — plus account auto-rotation as the safety net. The mobile vibe-coding setup that currently runs smoothest for me: how to build it, and the traps I hit.

模型越强，你说不清楚想要什么的代价越大。从地图与疆域讲到四种空白，拆一套我自己验证过的和模型打交道的方法，最后讲怎么让每个坑只交一次学费。

An agent's most dangerous output isn't bad code — it's the word "done." The whole point of a pure core, a one-command verify, a separate verifier, and an evidence ledger is to take the verdict out of the author's hands and make "done" a thing that's proven rather than felt. This is the substrate the rest of the series was building toward.

The reflex when an agent misbehaves is to add a sentence to CLAUDE.md. It's wrong in two directions: instruction files are context, not enforcement, so the rule won't bind under pressure — and the longer the file gets, the more it pollutes the context that makes the agent reliable. Load-bearing rules belong in hooks, lint, and CI. The prose should get shorter.

An agent is excellent at honoring a contract it can see and terrible at preserving one it can't. The boundaries that survive the agentic era are the ones written into the repo as typed, tested, importable things — and the ones that break are the stringly-typed promises that lived only in a team's shared memory.

The maintainer of your repository quietly changed from a human who remembers to an agent that re-derives everything from the files each session. Once you accept that, the repository stops being documentation and becomes an interface. Directory layout is the navigation API, contracts are the constraint API, and verification is the validation API.

Agent 最危险的产出不是烂代码，是"我搞定了"这几个字。纯核心、一条命令的验证、一个独立的判官、一份证据账本，全部加起来就为了一件事：把"做完"的裁决权从作者手里拿走，让它变成被证明出来的，而不是被感觉出来的。这是前三篇一直在铺垫的底座。

Agent 干错事，大多数人的反射是往 CLAUDE.md 加一句话。它错在两个方向：指令文件是 context 不是强制，规则在压力下根本不绑得住；而文件越长，越污染那个让 agent 可靠的 context 本身。承重的规则该进 hook、lint、CI，散文应该越写越短。

Agent 能完美遵守一条它看得见的契约，也能毫无知觉地破坏一条它看不见的。能熬过 agent 时代的边界，是那些写进仓库、有类型、有测试、能被 import 的；会被弄坏的，是那些只活在团队记忆里的裸字符串约定。把每条边界都变成契约，是手里杠杆最高的一招。

这几个月改动基本全是 agent 写的：人出判断和认知，执行层全是 AI。代码的读者换人之后，仓库就变成了 agent 要对着编程的接口——tribal knowledge 记录下来就能像软件一样直接分发，目录是导航 API，契约是约束 API，验证是校验 API。

I wanted a sushi-counter field guide that used the names you actually order by in a US restaurant — not Latin binomials. So I pointed a team of agents at it: parallel research workflows verifying English menu names with web search, gpt-image-2 for ~150 photoreal neta shots, and a deterministic HTML→PDF assembler. 73 fish, 92 pages, two debugging war stories, and a free download.

朋友给我一份用拉丁学名的 omakase 认鱼课，可板前点菜时，菜单写的是 Hamachi、是 Sea Bream。我想要一个标美式菜单常用名的版本，于是指挥一队 agent 搭了出来：九个并行调研 agent 上网核实菜单名、gpt-image-2 出约 150 张写实图、一段确定性代码拼成 PDF。73 种鱼、92 页，外带两个踩坑故事和一份免费下载。

Reactions to a podcast on AI-first org architecture. Use AI as a tool and you're racing the clock — 10x at best. Use it as a system and you rebuild the whole workflow, retreating to just two jobs: give intent, review output. But the review gate is drowning, and that's the part nobody wants to say out loud.