👋Hey Alex

Turns out AI Agents gets lazy when I do.

@mitsuhiko 3 years ago I created a t-shirt "Never say thank you to an AI". In reaction to people being "polite" with their chat bot. And now this? It's both hilarious and scary.

More musings after some people got upset about the word clanker. lucumr.pocoo.org/2026/5/26/clan…

@Everlier @badlogicgames We've been using these borders since forever in @WeStormz. I don't know of a better (in term of UX not fancy styling) way to color code cards. Instantly recognizable by a user, ultra visible. Won't change something that work because some people on Twitter are bashing the pattern.

@badlogicgames The em-dash of UIs created by LLMs, second worst offender is a badge with the status dot :) x.com/Everlier/statu…

if we see this, we know

@mitsuhiko The only thing I need is being able to check which skills or files are read. And if they are read in their entirety. Not their content. On the contrary I wanna see every single line of the thinking.

BTW if you don't like the new behavior: the read tool is swappable! But I would love to know if you want the old back why :) x.com/badlogicgames/…

@championswimmer yes, very interesting, will test it tomorrow. The default mode seems good enough for me.

Running a Pi session where current context would have. become ~400k tokens (i.e. would have hit compaction for GPT 5.4) but due to tool call pruning is at 133k (33% only). And new tasks at this stage are still being done pretty fine. Thanks to : pi-context-prune

@AmelieTabatta @Zellij_dev Same here. I'm in love.

I've recently discovered @Zellij_dev and it's such a delight to use. Basically tmux but it doesn't feel like a tool only appreciated by vim users that pride themselves in never touching their mice

@thdxr This is similar to iterating slices vs building layer after layer

finally found the right metaphor for this shift in how i use opencode. i used to treat it like 3D printing, where you build the thing layer by layer and commit to each piece as you go now it feels more like progressive rendering, you start with a blurry version of the whole thing, then keep making full passes over it, and each pass sharpens the entire shape doing this with gpt 5.5 and voice prompting is the first time things feel like they're clicking

@jessmartin @badlogicgames Thx, will have a look. At least the waterfall I've seen in the industries I worked with is one phase (and not slices) after the other, with formal gate reviews. And going back to a previous phase was considered a failure.

@heyalexnet @badlogicgames you should check out Royce's paper on it. lots of feedback loops! and lots to try.

it is absolutely crazy to me that our entire industry has succumbed to hyper waterfall. because that's what ya'll are doing with your massive plans and beads and dark factories. have you learned nothing?

@jessmartin @badlogicgames Waterfall with a feedback loop is just not waterfall.

@badlogicgames (I love pi) The reason waterfall was bad was because of long feedback loops. If you can run an entire "waterfall" process in minutes, the feedback loop is short enough the process can be effective. We should try waterfall again! Adapt it to the new constraints of production.

@mattpocockuk As usual, "it depends". There is no universal truth about how much planning you need for an optimal results

The more I replace plans with prototypes, the better the outputs Who'd have thought that low fidelity prototypes were better than walls of spec Oh yeah, the entire industry for 20 years Stop going against decades of knowledge because someone in SF shipped it as a 'mode'

@GregKamradt I’ve vibe-coded a lot of my internal tools. Not just because it’s cheaper, but because small custom tools just fit me and my workflows better. They do fewer things I don’t need. And more of the specific things I actually need.

Yes, you could vibe code all your tools But do you really want to? Or would you rather a professional team extends and maintains them for you (and your agents)?

@badlogicgames @rachelnabors @rivet_dev I hate it. It's beautiful but it's blocking access to the links just enough time to frustrate me 🤣.

@rachelnabors @rivet_dev i'm sorry.

agentOS by @rivet_dev is nifty, and not just because it's built on pi.dev. check it out! rivet.dev/agent-os/

@tombielecki @nicopreme When the next task does need the full context to get the job done right.

@heyalexnet @nicopreme Are there particular tasks/problems where this doesn't work?

addicted to using boomerang mode (aka reverse D-Mail) in Pi these days. ctrl+alt+b enables it for the next prompt submitted -> after the prompt runs, it rewinds back to the same point with file changes intact + leaves a summary in the feed so the agent knows what happened. using it often can make the context window feel nearly unlimited. Powered by the native /tree functionality in pi. pi install pi-boomerang
github.com/nicobailon/pi-…

@badlogicgames It doesn't 💪

@heyalexnet nope, try it. scream if it breaks.

People of pi.dev. pi has supported OpenAI WebSockets mode since February. However, we did not support delta updates. pi v0.71.1 now also supports delta updates: only the latest context additions get send, leading to a nice 66% throughput increase. Start pi, /settings -> transport -> websocket-cached Here's SSE vs. cached WebSocket side by side. This will also apply to your OpenClaw instance, should @steipete & team decide to update to the latest and greatest, giving you all the benefits of the pi runtime on top of the Codex server backend.

@badlogicgames Speaking of tokens, I have the intuition 5.5 generate more tokens than 5.4. About 10% more with a similar work process on pi.

compare tokens to requests. what does that tell you?

@startupideaspod Honestly , I can achieve the same level of performance with both tools. Process has way more impact than the choice of tools.

Codex is the new "super app" of AI tools. Here's the simple breakdown: Vibe coding + knowledge work in one place: - Cursor: code only. Can't make a doc you just view. - Claude Code: code only. Docs live in Cowork (separate, restricted). - Codex: both, same interface. Codex on complex tasks: - Better at infrastructure - Better at hard, multi-step problems - We tested it across our team The proof: - One-shotted a Replit clone. - Sandboxes. Mobile vibe coding. - 1 hour 20 minutes on 5.4. "The fastest way to do the most amount of things." That's why Codex is winning.

@thsottiaux @badlogicgames I don't assume a specific cause but there was something. Codex was unusable via Pi today (yesterday same). Tons of errors, significantly slower. On Codex Cli not a single error.

@badlogicgames Not true. Please don't spread misinformation out there, it's already hard enough to keep facts straight! Will investigate today if there are differences but the requests should all go to the same clusters irrespective of the client (pi, opencode, openclaw, w/e).

@mitsuhiko @ir_foo Also not everyone is using npm

@ir_foo Would you want it to update behind the scenes and then take effect on restart?

If something really bothers you about pi, give us feedback please :)

@badlogicgames Quite pi-ttoresque

i'm honestly surprised by the reactions to the announcement. twitter: 300+, all positive, with the Tolkien name being the only point of contention (see blog for my thoughts on that) bsky: my austrian possy, all positive, despite a general anti-AI stance, specifically from a societal perspective, which i partially share orange site: we don't talk about garry's site emotionally taxing. thanks all of ya for being so nice!