New blog post, wherein I beat a dead horse for the last time. https://t.co/j64CpIysH1

What I'm saying is: Bash is all you need. That said, you will soon be able to add your own tools to pi, if like a more integrated approach. pi itself will not offer subagents out of the box, because I think they are an anti-pattern (don't @ me).

Here's one way to do subagents in pi. General principle is always the same: let your main agent call into pi --print, with whatever arguments you want (specific model, thinking level, tmux for backgrounding, etc. pp). Skill: https://t.co/pabwti63I6

Neat idea!

Thought it would be fun to be able to summarize what I browsed last week? Or what sites I visit at lunchtime. Or to link me to all the blog post about css I read on the weekend. This CLI + Skill can help:

My process for having my agent talk to external apps/services: 1/ Check they have a CLI - if yes skip to step 4 2/ Check they have a API - if yes, downloads docs 3/ Ask AI to build me a CLI for the API with just the functionality I need 4/ Create a skill that calls the CLI when

There you have it. Infinite slop machine finally achieved!

our team is small, older than most and we never fool ourselves into thinking having 30 agents going is somehow productive

🔉 Introducing SAM Audio, the first unified model that isolates any sound from complex audio mixtures using text, visual, or span prompts. We’re sharing SAM Audio with the community, along with a perception encoder model, benchmarks and research papers, to empower others to

If you are one of the ones who cares about context management and wants to have tools and a clean context window, here is me blabbing about tool search: https://t.co/QOiUXSrKXb

Well, I suppose that's some data to work with. Kind of. Anthropic lowers the max image dimensions after 20 images in a session. Others may have similar weird restrictions. Sigh. Wonder if just scaling everything down to 2000px and only keeping the last X images in context is

I'm looking for a cross-platform (Windows, Linux, macOS), self-contained (read no daemon, etc.) way to run a small containerized Linux, without many bells and whistles. Must support networking (in/out) and mounting dirs from the host. container2wasm comes close. Anything else?

👀

Trying to figure out image size and count limitations across providers and models. It's funny how badly this is documented. I'd assume OpenAI (500 images max) does pruning in the background?

I remember running full blown operating systems with GUI that took less than that. I guess that dates me ...

Probably doesn't matter much in today's age, unless you run dozens of these, but there's something to be said about simple CLI apps taking multiple hundred's of megabytes of RAM. Except Codex. All empty sessions as well (except pi, +600 messages)

@badlogicgames I made a SQLite cli which you can use for everything: tasks, agent mail, memory etc.

Latest pi version does tool result streaming. Because I like watching Bash commands executing. https://t.co/q7nmrdJRDk

https://t.co/1OHn2HnXSK

Please send Thariq any examples of odd behaviour! We don’t nerf models - but reports can help us track down bugs.

One of the principles behind Amp always has been: what if you give these models enough tokens? what if you don't try to match it to a $20/month subscription? these models yearn for the tokens! Since then, in Feb/March, we've seen the $20/month basically disappear from the