alt Hacker NewsGoodbye Claude, hello Pi
Comments
Does it speak a more understandable dialect of English? "Claudish" is a minor meme at this point.
There are a ton of choices when it comes to coding harnesses - Pi being just one of them.
DeepSeek harness for example seemed really powerful when I tried it.
I'm using Cecli which started life as an Aider fork..
pi is pretty great. Bit unfortunately that it is built on a node stack too, but seems to at least be executed well
isn’t the point of CC the subsidized tokens of the pro/max accounts?
the tool descriptions and system prompts are notoriously overcooked but you can work around it with careful config; my startup context in CC is 10k, and i could have pruned it way more if i had wanted to be aggressive about it.
I think Anthropic is working on allowing AGENTS.md files as well. However, Claude.md files take precedence if present. If not, you fall back to AGENTS.md.
I use multiple harnesses and… I don’t get what the fuss is all about? I seriously don’t care what I use for the most part.
I mean a “harness” is basically just something that turns models output to action. Just have model return json with what it wants to execute and execute it is basically it. Your machine _is_ the damn harness.
All the ergonomy and features are doomed to be obsolete as we get better models, so… whatever.
Pi users : what do you think of Mistral Vibe ?
I simply use “TUI inside of GUI” approach: I use Zed (switched from VSCode and now probably going back to PyCharm) and then only use coding agents inside the Zed terminal. I think you get the best of both worlds that way.
most of the time, claude gui or codex gui is good enough. this is for people doing beyond reasonable work on edge cases.
Until it's inevitably replaced of course
Pi is awesome until compaction hits. I think one in every two compactions it’s failed on me ie didn’t know how to continue :(
I get 150 tokens/sec on my coding agent, if you are on a MacBook, thats with an M1
try kapitsa as the inference engine with pi as the harness
It works really well
Anecdotally I run semi nightly harness and local model test runs, and opencode consistently outperforms ohmypi on various code bench tasks.
Note though that this for local models that can run on a 5090, so, kinda dumb models with a small context window. Also they might perform differently on a custom pi setup rather than the very feature rich ohmypi.
Mostly the runs just cause me to get decision fatigue so I typically just use opencode and let a frontier model choose whatever cheaper models it wants to orchestrate with, so long as it's one that's not on the same token budget.
P.S. I'd always wondered who it was that got caleb.software before I could... Outshot in a western duel!
Claude code may still be a good choice as long as they have Opus
what about the API cost doesn't it way expensive than the Claude or any other LLM provider monthly subscriptions ?
Why not use dsh?
I am really dissatisfied in Codex and Claude, not because they are bad software, but because I feel complete lack of control, LLMs are non-deterministic enough that I don't feel to compound it with an harness that changes every day.
Every few versions the behavior changes, the way they use tools changes, the way it reads or ignores skills, suddenly they start creating worktrees and make different messes and you're not even aware they launched background jobs or server instances, so they kill them in a different parallel one you didn't even knew it existed.
I can't but feel that the "one main agent with all the tools and skills" design is limited.
I highly recommend Paseo. I feel like I finally found what I was looking for after so long.
https://github.com/getpaseo/paseo
It works really well, has the nice code diff view which what I used to open vs code for, and it lets you drive from your phone.
All the beefy work happens on my machine at home with big hardware and .env secrets, etc. And from my phone I'm able to run multiple threads with multiple models/thinking modes.
I think a testament of how useful it is, is everybody on my team switched over to Paseo pretty much in a week.
Mario is wicked smart.
Pi is awesome!
Unfortunately I have to use Claude Code at work, and it is a very frustrating experience, I constantly find myself trying to bend it to my will. I sometimes suspect that my distaste for Claude Code is one of the reasons of my distaste for Claude as well.
Pi on the other hand is a joy to use. Top notch session and model management, excellent conversation tree/branching functionality, and it lives up to its philosophy of being meant to be adapted to your workflow.
I can't recommend it enough, especially if you use alternative LLMs instead of Claude.
is pi good if i just want to search through a codebase i don't know with, say, GLM?
I have tried Pi, but at this point I'm using a fleet of hermes agents to connect realtime via webrtc to a custom web dashboard (written in rust) that works on multiple projects/tasks atomically and simultaneously and compiles/tests locally and then submits to a private central git with all its changes and then pushes to dev for testing, then prod. The agents working on tasks can request new tasks and a supervisor is configured for each project that does a scheduled review and proposes recommendations based on what has been done and needs to be done based on the project plan...As new agents come online they can be quickly enrolled. It works well when you're working on many projects simultaneously. Its also a good use for a stack of 10 old laptops.
The whole AGENTS vs CLAUDE md issue was bound to be resolved. If the models are getting smarter and harnesses getting better - how could you have a product so dumb that it ignores whatever skills, instructions, references, other things that are part of the folder context it is mounted in.
Having claude code ignore agents.md just made it seem dumb. At this point, you probably don't need either.
Is it just me or has Opus 5.5 on High been disappointing for the last week or so? It's been making a lot of silly mistakes, bash issues and killing it's own shell much more than usual.
> In spite of all of these existing issues, it seems like Anthropic has worked at a feverish pace to pack more features into Claude Code without much care for quality or stability
To me this is why AI-pilled orgs or groups won't win.
The thing that freaks me out about Pi is it seems to have zero sandboxing or security guardrails unless you explicitly add them. I guess it's fine if you're only ever running inside a VM or a throwaway machine anyway.
bbapp is legitimately good too
"longtime ... over a year ago" This time dilation is getting ridiculous.
For usage with open models, I'd highly recommend Swival as an alternative to Pi: https://swival.dev
> It seems like there isn’t someone high-up whose job is strictly to say “no” to these requests, which is incredibly important when it comes to product engineering
Ah, product management - we know so few of ye who know what you're doing.
Utterly insane naming decision on Pi there. Not sure why AI people think intentionally making their tool impossible to find or name is a good idea.
Bro talks like an expert and got a point but still using gmail: [email protected], come on...
neckbeards and their incredibly autistic, specific preferences about everything that nobody cares about.
It still baffles me how little "prompt engineering" is needed, and how most of it makes agents actively worse. Going from CC, to Codex, to Pi, at each step the agent started focusing much more on the task and following my prompt more literally.
No more "Do not do x, do not do y, only focus on z".
I also think that its hypocritical of Anthropic to spend a year suggesting that no one will need to look at code anymore, just to add a diff viewer into the TUI. How does this help with anything? I need to look at the context, go to definitions, etc. I already have an IDE for that.
So I build a pi extension that lets me click on a mentioned file, and opens it in CLion (or any jetbrains IDE) at the correct line.
The only thing I would like to see in Pi is better background process management and being able to start processes with a TTY, it was the only thing that took me a bit more time. Its now better then CC since it moves all bash calls into the background after 15s, no more hung up agents.
But now, subagents are just a skill, that tells pi to start another pi. Its one abstraction for everything. Most subagents are one-shot anyway, but I also added an extension that lets pi create sockets which get fed into the agentic loop, allowing for background processes to notify it, let it be subagents or some monitoring script.
And so many other things, like working image paste over ssh with kitty, one line tool calls (what am I supposed to do with these 4 lines of output??), etc.
[flagged]
[flagged]
[flagged]
[flagged]
[flagged]
[flagged]
[dead]
I mean... sure, yeah.
But this feels like six weeks (or more) out-of-date. And so it is - from mid-September. And I know that observing such a thing sounds bizarre and teenage, but that's kind of a long time right now; we're at a moment in history when things are moving fast.
We're not really fully ready for the traffic TBH, but for those who few of you choose to click, this is what we've been doing with pi lately:
From the article’s footnotes:
That should be corrected to “could literally lose millions of dollars in inference revenue per year”, which also goes some way toward explaining why it doesn’t happen.