Joe Crowley

Joie de vivre. Thermalizing.

Physics: Cal Poly ↦ UC Santa Barbara ↦ Fat Tailed Solutions.

[Joe Crowley @crowleylab]

incredibly frustrating when an xhigh model replies without thinking first, or doing any tool calls.

[Joe Crowley @crowleylab]

switched to opus 4.8

[Joe Crowley @crowleylab]

why wouldn’t anthropic call it a “fable” and deliver a short, fictional story crafted to teach a moral lesson?

[Joe Crowley @crowleylab]

why wouldn’t anthropic call it a “fable” and deliver a short, fictional story crafted to teach a moral lesson?

[Joe Crowley @crowleylab]

yes get it all integrated and complete please... yes approved completely, and get all the remaining parts done and finish it in this pass, continue until it's done

[Joe Crowley @crowleylab]

@TakoTreba loops are bad
agent heirarchy is better
agents spawn subagents spawn subagents .... until the task is nearly trivial per agent.

[Joe Crowley @crowleylab]

@housecor loops are bad

agent heirarchy is better

agents spawn subagents spawn subagents .... until the task is nearly trivial per agent.

[Joe Crowley @crowleylab]

@tekbog have codex mine your data and RPA you

there's literally thousands of hours of it already, right?

[Joe Crowley @crowleylab]

@yacineMTB this is the problem with talking too frequently with llms, seemingly smart but nothing behind the curtain

smart people are worth talking to you because they steer you too

[Joe Crowley @crowleylab]

loops are the wrong basis for the capabilities here

use heirarchical delegation with parallelism + phased sequencing instead

loops get too nested. figure this out before you spend all your time building loops.

[Joe Crowley @crowleylab]

@beffjezos @steipete there's a phase transition as token limits expand

human input tokens saturate

how do you run more agents? a layer of abstraction.

agents run agents. how do they run more agents? a layer of abstraction.

turtles...

[Joe Crowley @crowleylab]

we're the ones building the loops because they have not been achieved internally in a generalizable way people want to use the ai to do the stuff they care about not spend their tokens building the system to spend more tokens i would prefer it just be curious about me, ask me questions proactively, and work on support tasks that help me without me prompting for it.

[Joe Crowley @crowleylab]

@sama the 10x was really sweet while it lasted. i was able to utilize nearly all of it, including resets. it taught me how to apply agent-to-agent delegation well. and it showed me how compute hungry the world will be.

[Joe Crowley @crowleylab]

@thsottiaux the 10x was really sweet while it lasted. i was able to utilize almost all of it. it taught me how to apply agent-to-agent delegation well. and it showed me how token hungry the world will be.

[Joe Crowley @crowleylab]

we're the ones building the loops because they have not been achieved internally in a generalizable way

people want to use the ai to do the stuff they care about

not spend their tokens building the system to spend more tokens

i would prefer it just be curious about me, ask me questions proactively, and work on support tasks that help me without me prompting for it.

[Rahul @sairahul1]

Boris Cherny, Head of Claude Code at Anthropic:

"The next transition is coming this year."

2023: you wrote the code.
2024: you prompted Claude to write the code.
2025: you wrote loops that prompted Claude.
2026: you build the harness that runs the loops.

Each transition, the engineers who understood the new abstraction first moved ahead permanently.

The ones who waited caught up eventually — but never all the way.

Harness engineering is the 2026 abstraction.

Here is everything you need to know ↓

[Joe Crowley @crowleylab]

frederick taylor is a total menace for humans

he would have done pretty well directing a swarm of agents

[Joe Crowley @crowleylab]

@lanyi1992 "the limits of my language mean the limits of my world"

LW

[Joe Crowley @crowleylab]

imagine a scene in your mind

describe that as articulately and expressively as you can

feed that as a prompt to an image gen ai

how different is your mental image from the genai image?

it's my frustration using codex. it's most thinking too, not limited to images.

[Joe Crowley @crowleylab]

wittgenstein
"the limits of my language mean the limits of my world"

[Joe Crowley @crowleylab]

imagine a scene in your mind

describe that as articulately and expressively as you can

feed that as a prompt to an image gen ai

how different is your mental image from the genai image?

it's my frustration using codex. it's most thinking too, not limited to images.

[Joe Crowley @crowleylab]

where do i upgrade from $200 plan to?

[Tyler @rezoundous]

69% chance I'll be forced to upgrade Codex from $100 to $200 plan.

[Joe Crowley @crowleylab]

really feelin it today...

[Joe Crowley @crowleylab]

unfortunately with the 10x * 2x usage limits

i've gotten used to it

and i'll spend my last week with it

building the much leaner agent swarm

that repeatably does the workflows i used

with 5% the token budget