I agree with the author on a lot points about the joys of using low cost models. My work would pay much more, but I prefer to use cheap models usually. Assuming cost correlates somewhat closely with energy used, I feel good about spending as little as possible, and the cheap models are so capable.
I feel the opposite. An agent sometimes work an entire day . The chance of doing that work twice defeats any motivation to use cheaper , less capable models . Also if you open benchmarks leaderboard for the first time in a while there’s almost 0 Chinese models .
Nothing says that agents can't read skills, interpret them and fetch stuff. Needing to agent to pass things off to powerful LLMs to interpret doesn't have to be done for everything.
Yes, pi's works a bit better since it exposes the whole tree to you, but in the simple case of forking to a new conversation it is about the same as Cursor and probably other coding agents.
One thing you can do in pi that you can't in Cursor: Have a 5-prompt conversation, jump back to prompt 3 and have a new conversation [call this convo2], then jump back to the original point 5, then jump back to convo2.
So fork is really only needed for when you need to interact with the conversation tree in two separate processes.
I disagree on plan and todo-list, which seem relatively not needed to me. No comment on goals/looping, which I don't have much experience with.
To me it's great how minimal the system prompt and tool set is, and I doubt those features are worth the tokens for every model. (Who knows if they improve performance for SOTA models, and they probably harm performance for small local models.)
I would add "better sandbox" support, which I think should be included out of the box. Not having a very simple way to get out of "yolo" mode is kinda crazy. Sure, there are plugins, but they do have some issues.
pi-bash-approval is good, but manual approval plus allowlist is a "bad" way to run coding agents.
The "best" way in my current opinion, is where commands run by default in a sandbox, but commands can be ran unsandboxed as needed, requiring approval or allow-list in that case. Cursor was pretty good at this (when I used it). For example I don't want to configure my sandbox with access to docker, which would present easy jailbreaks, but I do want to allowlist certain docker commands or approve them to run on my host as needed. pi-sandbox is good at allowing me to configure sandbox access, but this feature where somethings can run unsandboxed is missing. I write more about this and about (hacky, but productive enough) workaround here: https://github.com/carderne/pi-sandbox/issues/50
It starts up for me in under a second on my M4 mac. (Obviously could be faster, but doesn't bother me personally on my hardware)
My one main "issue" is with using the pi-sandbox extension. It's based on a forked claude code sandbox runtime. Not exactly sure why a fork was needed, and the fork is a bit behind now. I also wish the sandbox feature worked a bit more like how Cursor's sandboxing worked. Not familiar with Claude Code sandboxing, so can't compare that. I describe the issue and a (slightly hacky, but productive enough) workaround here: https://github.com/carderne/pi-sandbox/issues/50
Obviously the fact that I can fork a plugin for pi and customize it as needed is quite a plus too. Really all the other features work quite well for me!
