Didn't anyone think about making an agent harness with no access to a shell? Just read files, edit files, list, search and so on?

Giffen good: someone has a fixed agent budget they spend partly on Claude and partly on Deepseek. As Deepseek prices rise, they're forced to use Deepseek more and Claude less because Deepseek is still cheaper and the total amount of tokens they need won't get small.

I almost got it right, as this was released:

fx

fx - Tiny, open, native coding agent

Tiny, open, native coding agent. A minimal CLI written in Zig, optimized for research and embeddability as part of larger systems.

> fx is a coding agent harness and CLI written in Zig, optimized for research and embeddability as part of larger systems. View quoted note →

In Europe in the 19th century sending letters was very expensive. But it was charged by letter, not by weight or size of the letter, so people would gather to send multiple letters in the same paper or envelope, and at the destination it would be split and distributed to the actual recipients.

I mentioned here that "pi can build itself", because that's some of the marketing I hear. I've heard similar claims from other competitors. But in fact I never liked that some people refer to the software that runs a glorified chatbox to an LLM with access to files on disk an "agent" and then claim that it somehow acts when in fact it's either the human prompting it that acts or, at best, the LLM model itself that "acts". View quoted note →

pi.dev has some weird marketing. It advertises itself as very simple because it only ships with 4 tools: - read - write - edit - bash But it's weird, because read, write and edit can all be done from bash directly (and the models will gladly do it if you try to block any of the others), so why not only say it ships with 'bash'?

LLM models think, just like a rock also thinks. The rock, though, is just a very dull guy who just wants to sit there and do nothing all day. Except when it rolls, in that case the rock decides it likes to keep rolling for a while until it stops. LLMs must have different thoughts, but not too different.

hunk

hunk — review-first terminal diff viewer

Hunk is a review-first terminal diff viewer for agent-authored changesets. Multi-file review stream, inline AI annotations, watch mode, TypeScript ...

is a good tool to have in your toolbox.

I've watched a guy talking about how Pi was the best coding agent in town and it could be customized and it was so much faster. I tried it and basically spent a day trying to learn how to customize it, only to give up in the end. Turns out it isn't faster than OpenCode, that was a lie. It is also not great at customizing itself, it has to read all its own docs all the time. I guess I'm just not used to letting the vibe flow and handle control of all my computer to a remote access trojan.