saturatedfat

pushshift baby

I think for days at a time still.

I don’t think you can get the same satisfaction out of these tools if what you want to do is not novel.

If you are exploring the space of possibilities for which there are no clear solutions, then you have to think hard. Take on wildly more ambitious projects. Try to do something you don’t think you can do. And work with them to get there.

let me take a shot. i have thought about both for a while.

dspy is declarative. you say what you want.

dspy says “if you can say what you want in my format, I will let you extract as much value from current LLMs as possible” with its inference strategies (RLM, COT; “modules”) and optimizers (GEPA).

gas town is … given a plan, i will wrangle agents to complete the plan. you may specify workflows (protomolecules/molecules) that will be repeatedly executed.

the control flow is good about capturing delegation. the mayor writes plans, and polecats do the work. you could represent gas town as a dspy program in a while loop, where each polecat loops until its hooked work is done. when work is finished, its sent to the merge queue and integrated.

gas town uses mostly ephemeral agents as the units for doing work .

you could in theory write gas town with dspy . the execution layer is just an abstraction . gas town operates on beads as state . you could funnel these beads thru a dspy program as well.

the parallels imo are mostly just structured orchestration .

i hope this comes off as sane. 2026 will be a fun year.

i made a clone of beads in rust that uses CRDTs for real time sync to coordinate a bunch of coding agents at the same time! if running locally, it's instant, for git it takes around .6s after your last action. lives entirely in git, and is like... an actual distributed database/issue system, just works, and u never gotta think about it.

https://github.com/delightful-ai/beads-rs/tree/main

it doesnt control for skill using models/experience using models. this looks VERY different at hour 1000 and hour 5000 than hour 100.

unfair characterization.

think about how this thing is interacting with your codebase. it can read one file at a time. sections of files.

in this UX, is it ergonomic to go hunting for patterns and conventions? if u have to linearly process every single thing u look at every time you do something, how are you supposed to have “peripheral vision”? if you have amnesia, how do you continue to do good work in a codebase given you’re a skilled engineer?

it is different from you. that is OK. it doesn’t mean its stupid. it means it needs different accomodations to perform as well as you do. accomodations IRL exist for a reason, different people work differently and have different strengths and weaknesses. just like humans, you get the most out of them if you meet and work with them from where they’re at.

could u share some more intuition as to why you started believing that? are there ANY comments that are useful?

for the love of the mf game. most fun i’ve had with computers in my whole life

very very cool .

yes but ud ideally want ur own high uptime mac mini

heya, building this. been used in prod for a month now, has saved my customer’s ass while building general workflow automation agents. happy to chat if ur interested.

darin@mcptesting.com

(gist: evals as a service)

writing an essay on this if you'd be interested in reading a rough draft. i believe this very strongly and i think it's the UX that got us here and it's UX that'll take us out.

quintessential hacker news comment. thanks for this.

Would you guys consider adding a llms.txt ?

Is the OneTool source not available?

Also, do you guys do something similar to: https://github.com/frankfralick/dripgrep ?

(for the curious, the README for this repo is 100% worth taking a look at!)

Really cool!

Had a couple questions:

- how far does the $50/mo get you?

- what's the battery life like/going to be like?

- do yall allow people or plan to allow people to swap out (some) of the models or orchestrate them yourself, at least incrementally? say I want to fine-tune my own maurice personality

why can’t you feed the chat history back in at every forward call?

wrote a blog post on using DSPy and why it’s powerful. tried to keep simple, understandable, concrete.

https://www.darinkishore.com/posts/mcp#building-tools-that-l...

hi! wrote a lil blog post on tool design for MCP, was wondering what you think?

https://www.darinkishore.com/posts/mcp/

Supermaven, but you don’t get model choice.

It’s helpful for people who are new to using things on the terminal. I use a similar version with the ctrl+l hotkey, you can do most things on the terminal just by typing english now.

That+fish is an amazing introduction to the shell. Totally new avenue for learning and being immersed quickly.