Nikita M.
@meln1k
Software engineer @ Datadog by day, coding agent hacker by night. Building
245 Following 1.3K Followers
celld is so good. compute providers have ZERO vendor lock-in on me now.
amp became THE control panel for all compute I own:
gaming pc, diy NAS, macbooks, chinese mini pc, even some old hardware from 2016 is brought back from the attic and now hosting the amp runners. even a steam deck has an amp runner on it.
Show more
remix 3 + effect is a surprisingly good combination. just those two libraries, and you almost never need to tap into anything else for your web project.
most of the time deepswe bench is judged on the cost, but the agent steps per task is more interesting because it shows how much slop you will get. my day to day experience confirms this too.
Show more
Get exposed to the problem before the agent. Make your first contact with reality agent-free.
Don't let the clanker corrupt your original thinking. Struggle with the problem, synthesize the solution, only then ask the bot for feedback.
Steer the agent, don't be steered by it.
Show more
yall im ngl its way harder to get joy and satisfaction out of building with ai than it was before
constantly straddling burn out, being far less immersed in hard problems, constant context switching
i’m tired of the uncertainty of where this is going and how to do it well
Show more
if you’re hosting your projects on a VPS instance and not a PaaS, how do you do that? SSH into a box and let the AI agent manage it? DIY setup with docker compose? dokku/coolify?
before agents: a problem pops up, you learn half of company repos, build understanding, fight the dragon, and return enlightened.
after agents: you send the clanker to figure things out, finish the task in 2 hours but learn 1/10 of the traditional way.
is it worth it?
Show more
another weekend spent trying agent loops. no success for high quality code gen so far. The best results I've got is by reducing the amount of LLM output I need to review:
1. review at the leverage point, aka plan.md
2. use linters/tests/strong type systems to reduce the llm output entropy
Loop people, how do you run agents unattended and get quality results?
Show more
got a powerbank with replaceable battery + two pouches of 20Wh 21700 cells, should be enough for 9 days of uninterrupted agentic coding from phone
