Geoff Goodman

@J_Kreutzbender @RhysSullivan @maxleiter It would need to become a convention adopted by agentic harnesses acting as MCP Clients. Just like ChatGPT Apps happen to use resources in a particular way.

@RhysSullivan @maxleiter My understanding is LLMs can't invoke resources like they can invoke tools? The MCP spec says that resources must be user opt in? It 100% make sense that skills become resources. I've created a `get-skills` tool that loads skill markdown files to get around the issue

It's unfortunate that MCP actually solves pretty much all problems people have with agents today, it's just that all of the first implementations of it were bad so people discredit it now The latest one is skill distribution, works so well for that

@mitsuhiko It's like this sawtooth pattern. Two steps forward (vibing) and one back (manage entropy). Agents are definitely on the side of thermodynamics.

(And I'm talking purely from personal experience here)

The biggest issue for me with agents is that they are hard to resist. But then you can build yourself some shit into the codebase that you get to regret in record time. And no, I don't think you can vibe yourself back to sanity with better models.

@DavidKPiano A lot of this is how pi.dev works already, FWIW. I think you might find /tree to your liking.

Been thinking about this a lot (and prototyping). Rough thoughts on how I think conversations should work; brain-dump for sake of sharing: - Tree data structure, nodes have parent IDs - Threads can either trace full path to conversation root or summarize from "branch" point - Threads can have threads, inf recursive - Agent still sees linear conversation (path to root or summary) - Threads are first-class primitives with their own metadata (name, purpose, etc), not just implicit from tree - Threads can run in parallel: agents can do work in multiple threads simultaneously - Messages have sender identity, not roles: "user" / "assistant" is a two-player limitation IMO - Messages can reference other messages and threads by ID: agents can cite, link, and build on each other's work - Events are the source of truth (very actor-model coded); messages and threads are derived views - Tool calls for creating, reading, and summarizing threads - Compatible across providers, of course --- I also have some ideas around multi-agent participation, state machine-driven agent behavior, & structured conversation flow that make this significantly more interesting

@badlogicgames "I like the interplay of colours with my theme." "This line sparks joy."

new rule: instead of attaching agent session logs to your prs nobody will ever read, have the submitter attach a video of them explaining every single line of code and why it must exist.

@ShrekOverflow They make you cry?

Sandboxes need to be like onions and not pineapples.

@mitsuhiko Which model?

I feel like I need to start fighting this shit with more linters. They are just not listening.

@mjackson @gdb I bet it's incredible for docs. Been struggling to get it to rein it in on design assumptions and to get it to push back on my nonsense like 5.3. It's an amazing general purpose model but I'm back to 5.3 for production code.

@filearts @gdb Yeah I’m doing everything with 5.4 on /fast. Making heavy use of skills. It’s writing way better docs than 5.3 ever did for me.

true

@mjackson @gdb Been struggling to get 5.4 to be as ruthless of an engineer as 5.3-codex. Haven't tried pro though which may be my issue. Have you moved entirely to 5.4?

@gdb Codex team cooked with the desktop app! I completely ditched the editor I’d been using for over a decade. 🤯

@filearts @badlogicgames @mitsuhiko @newsystems_ I'd happily come to Montreal for an impromptu event if someone puts it together (J'aime toute opportunité ou je peux pratiquez mon Francais)

@bentlegen @badlogicgames @mitsuhiko @newsystems_ We'd be two anglophones talking to each other in french for ... reasons. I too would attend but am not ready to organize.

🇨🇦 Incredible things happening in Toronto this Saturday - Live Q&A with Pi creator @badlogicgames and @mitsuhiko - Live demos from the city's most AI-pilled agentic coders - Free pie 🥧 - Hosted at New Stadium (@newsystems_)

@chris__sev Damn.

TIL that this emoji 😂 is literally called "the boomer emoji"

Anyone else feeling like GPT 5.4 is a regression in _real engineering_ so far compared to 5.3-codex? Personality is much more interesting but for writing production code, 5.4 is going too far, making bad assumptions and trying to be too helpful. 5.3-codex didn't try to please.

@kentcdodds @ShrekOverflow @ianhaz Probably not on Twitter. But agree that he's great. We have a few standout product people in-house. Interesting reflection of how there's this whole world of incredible people who just don't interact with Twitter.

@ShrekOverflow @ianhaz? Or someone with a very similar name!? 😅

Who is the very best product manager you know?

Me: "I need a USB-C cable" Codex: "Here is your USB-C cable"

@ShrekOverflow @yenkel @jfroma Lol @ name choice. Sounds like a recipe for disaster.

@yenkel Shallan (my OpenClaw) does not access Moltbook. She found the whole split personality thing a bit weird. (@jfroma I think you will get this reference)

this is not a social play. this is a data play. see it now?

@badlogicgames @steipete Agentic coding tends to entropy. It's like in the heyday of Apple dongle hell. AI will pile adapter upon adapter. You get the right outcome, but behind the scenes, entropy and code will inevitably trend upward.

your biggest enemy is still complexity. it's also your agent's biggest enemy. but it has no holistic view of your code base, so it keeps adding complexity. and you think that's how it's supposed to be, because the clanker shat it out, and you don't know the stack. glhf!

recommended reading. had this on my reading list for a while, finally found time. it's a small study, ymmv. would love to see one with senior engs. they great atrophy has begun. arxiv.org/abs/2601.20245

@badlogicgames Yes. Seeing the same. New today AFAICT

Anybody seeing this? Codex endpoint craps out, then recovers.

@evalstate @RhysSullivan Happen to have one for servers? 😅

@evalstate @RhysSullivan Oh wow. Great resource. Thanks @evalstate 🙌

The core of this system is MCP elicitation When a destructive action like `await tools.vercel.dns.removeRecord` is called, it triggers an elicitation from the client to approve it More harness should bring support for this one, is an incredibly useful primitive