Post

yeah, I was collecting antipatterns from a team who works with customers and started listing off what they do wrong like I had pushed exactly the right button!

amazing things are happening

I watched someone the other day consume a few thousand tokens rather than just cut and paste a command line the AI had already provided.

this is a process that’s slower to catch on than I’d like just the way you freeze essential tasks into dedicated silicon for better performance, like H264 decoding, a lot of agent stuff should be on the lookout for repetitive inference that can be reduced to cheap scripts

A big part of what I'm doing "at work" lately has been designing systems where the deterministic, non-deterministic, agent and human parts/interactions are mapped out. Oftentimes the non-deterministic (agent) portions get loops that gradually decreases their footprint as the architecture is tested.

gotchu fam!

unfortunately I will have to wait 72h for my usage to reset to run this string find-and-replace

There are limits where this behavior flips (money is a good example, below a certain point recklessness/gambling suddenly becomes relatively optimal, "because fuck it.") But generally, yes.

“Premature optimization is wasteful of resources” isn’t something I ever put info words, but I feel like I knew it to be true.

I find frontier models naturally kinda do this in Prime Agent, because the harness is just an IPython session.

"ok claude now do the same thing to the 50 other servers" 😬

Haha I think different folks mean very different things by tools ;) Think dedup when traveling on a graph. Do you give every agent a tool to update a singleton to coordinate within a swarm, or do you code a breadth first search calling agents on nodes? Same result, opposite architectures :)

Usually I'll blend in a comment that contains argument front matter so that I don't have to watch the agent try and fail to use jq 10x to figure out a commands args Mixing this in with this skill: x.com/steveruizok/... Has been noticeably better for my tool calls

just is so fucking good here. i want make to be as good, like, it feels like it should be, but it simply isn't, claudes in particular fucking love just

good models are supposed to tell you to do it the good way

I would say a huge portion of working with LLMs is scolding them for not writing scripts and making them go back and do it right.

I think this is going to be a generational divide. If you didn't tmux into servers or run ci jobs or write operators/controllers to manage infrastructure before ai agents, it's not going to be obvious to do that if you're coming into the industry now. Servers yearn to be pets.

me when i tell my agent to git push

Been doing a lot of latex, powerpoint and html-to-pdf that wouldnt get done at all without both sides of the tooling

Basically thinking (what to do next and whether to do it) can either sit between parametrized deterministic calls (tools), or itself be a call within a harness