最好的 AI Coding 工作流，可能是让 AI 把自己的不稳定性慢慢沉淀成系统。

文章作者用 Fable 开发面向 LLM 时代的源码管理系统，体验很真实：模型很聪明，能读大量代码、提 issue、修问题，但也会犯很低级的错，比如两次把 build/ 目录提交进项目。

LLM 会越来越强，但未必会完全可靠。它可能理解错约束，也可能在明确指令下继续做不该做的事。

所以 AI coding 的重点，不该只是让模型多写代码，而是把它经常重复的操作脚本化，把它容易犯错的地方做成自动检查，除此之外也要把它需要的上下文提前整理好。

模型负责探索和判断，工具链负责兜住它的不稳定。

最后真正有价值的是把这些经验沉淀成更可靠的开发流程。

https://replicated.live/blog/away

away

Source: https://replicated.live/blog/away https://replicated.live/## Automating away

A. Karpathy once said that OpenAI researchers are effectively “automating themselves away” by improving their AI. Right now I develop Beagle SCM with Anthropic’s Fable and it is of course a brilliant model able to spot nits in a mountain of code, file tickets, make fixes. Still, yesterday it managed to commit thebuild/dir into a project, twice. It is brilliant, but clumsy.

Due to the nature of LLMs, this issue is not going away as they progress further. They tend to be imprecise and non-deterministic.Ragelthe parser generator can “code” a 10 KLoCformallycorrect parser in an instant, deterministically. What about Claude? Well, my instructions say in all caps: DO NOT PARSE ANYTHING MANUALLY, EVER. It would be torturous and it would be faulty, just don’t. It tries anyway, so periodically I tell it to scan the codebase to find and remove any attempts at manual parsing. That mostly works.

It becomes ever more brilliant, no less clumsy.

The way to deal with an expensive, slow, clumsy but brilliant LLM is to give it fast, powerful and deterministic tools AND to build the entire thing into a deterministic formal workflow. Make it faster, make it see the relevant stuff at the right time, make it less clumsy, make it self-correct. Sandwich that brilliant but inconsistent non-determinism between powerful deterministic tools and equally formal processes.

This story becomes even more interesting if we make the tools and processes malleable. That way, if Claude does some sequence of actions too often, we automate it. If it fails at something repeatedly, we automate the verification step.

Essentially, we let the LLM automate itself away, in favor of simple reliable deterministic tools.

Beagle SCM lets LLMsscript their own routinesin JavaScript. While all theheavy liftingis implemented in C and rarely touched, the tooling layer (the lower part of the sandwich) and the workflow layer (the upper part) areall JavaScriptand pick their code from the filesystem,node\_modules-style. Imaginegithooks that can tokenize source files in almost any language, inspect file history and commit history, cross-check links, and basically reach any datagitcan reach internally. That is Beagle.

HN