John Munson

My employer shares all these opinions. Obviously.

[John Munson @curtismunson]

Bar for bar, feature for feature, an LLM still spits out far more code by default. Laziness is a virtue that the agents have yet to learn.

[signüll @signulll]

it’s clear we need to retire the term vibe coding because the models now produce better architecture & way more coherent code than an average software engineer.

perhaps even better than top software engineers too because unlike a human, ai has a much much larger context window.

this ain’t vibe coding at all. it’s precision engineering.

[John Munson @curtismunson]

@jeffrschneider Spec design looks seductive at first, but actually wicked hard. Especially distributed systems. This looks like an excellent direction. Congrats Jeff

[John Munson @curtismunson]

@badlogicgames @devagrawal09 @mitsuhiko Extending durable to codemode seems... hard. Wild if this is a property that just fell out of the design and just works.

[John Munson @curtismunson]

@badlogicgames Also, huge congrats on Pi 1.0 and Durable Pi. Absolutely love the direction. And I absolutely love the restraint and deliberation it took to get there.

[John Munson @curtismunson]

@badlogicgames just spit out my coffee 🤣

[John Munson @curtismunson]

@bcherny lol

[John Munson @curtismunson]

Finding Jev-shaped problems really is quite fun. Once you get the hang of it, you start to see them everywhere.

[John Munson @curtismunson]

@yacinelearning If that's what they're doing... isn't that just a form of backprop?

[John Munson @curtismunson]

@jeffrschneider @AgentMeshHQ Pretty solid set of skills. With the core workflow, you'll probably be using 3-5 of them 95% of the time. And if you're not using CC, you may find yourself making manual tweaks here and there. Overall though, definitely recommend.

[John Munson @curtismunson]

@NateSilver538 @micsolana predict next token + a few tokens on reflection + revise if needed

a bit more complicated than that, obviously, but conceptually that's all it is. all it took was figuring out how to get the models to "think" longer. i.e. how to let the models spend more compute on cognition.

[John Munson @curtismunson]

@rauchg A testament to how dumb people really are

[John Munson @curtismunson]

I thought to myself, maybe I'm spending too much time carefully designing agent workflows and constraining their trajectory. Maybe I'll just try out these fancy Claude Code dynamic workflows and let 'er rip. jfc. So wasteful. So dumb. Slop everywhere.

[John Munson @curtismunson]

@JayaGup10 Some people hear "routing" and think that we can just start making routing decisions at every single turn of the agent loop. Not true, destroys the LLM cache, actually more expensive.

[John Munson @curtismunson]

@mitsuhiko @juniorro16 @badlogicgames Good to know!

[John Munson @curtismunson]

@mitsuhiko @juniorro16 @badlogicgames Currently, it seems like the best bet is to fork all of the existing built-in tools and create codemode-friendly versions for each of them. That's definitely viable. Was just hoping to avoid that.

[John Munson @curtismunson]

@mitsuhiko @juniorro16 @badlogicgames My agent is saying the existing pi APIs don't really provide a straightforward path to implementing codemode and would require deeper gymnastics and monkeypatching.

@mitsuhiko @badlogicgames I'm sure you guys have studied this before. Does that generally seem correct to you?

[John Munson @curtismunson]

Of course, this code is completely ephemeral and only exists as part of the agent's internal cognitive machinery. i.e. codemode

[John Munson @curtismunson]

I'm excited to see what Jev can bring to the field of harness engineering and whether or not it will lead to the emergence of a new paradigm that can supersede ReAct.

[John Munson @curtismunson]

I don't think you can simply give an agent a Jev tool and expect to get good results. I mean, you could, but Jev is most powerful when embedded within code. So, let the agent tackle System One decisions at scale by writing Jev-infused code.