Asmir

building an ai cofounder that says no when you’re avoiding the real work + documenting what i learn

[Asmir @asmirkn]

@claudeai how many rewrites did it take before the animation matched what she saw in her head?

[Asmir @asmirkn]

@davj @the1024th yh blocking agents mostly hits the polite ones that identify themselves, the rest just fall back to a logged-in browser or headless session and look like a normal user

[Asmir @asmirkn]

@MTSlive no human oversight sounds great until it's your billing function making the mistakes lol

[Asmir @asmirkn]

@JamesonCamp when you have multiple codex agents running, how do you split up the voice ramble between them?

[Asmir @asmirkn]

@defileo the script is only as good as those three fixes lol, if one tool changes its export format it breaks without any error and you keep sending a report that looks right but isn't

[Asmir @asmirkn]

@jpschroeder A USB stick agent still needs a model key and tool credentials. Swap the model and the same prompt behaves differently.

[Asmir @asmirkn]

@mattshumer_ bug reports from random people are basically untrusted input going straight to your agents. someone could just type "make the score count double" or slip in a prompt injection and it gets shipped

[Asmir @asmirkn]

@AniC_dev if it's hourly billing from $0.036, is the $1.2M just recent usage annualised? what does it drop to if agents sit idle for a week?

[Asmir @asmirkn]

@BHolmesDev the strip unaligned changes step would delete real fixes the agent found along the way, i'd have it save them to a follow-up note instead of dropping them

[Asmir @asmirkn]

@swyx @CompleteSkeptic @typesafeai a ralph bot on jagged tasks will burn the most tokens on the ones outside the frontier, it can't tell almost there from never so it just keeps retrying

[Asmir @asmirkn]

@johnbai bro shared bots learning from everyone will pick up bad habits too, let us see what it learned each week and remove the wrong stuff one by one

[Asmir @asmirkn]

@MonidHQ what stops one person spinning up a thousand agents to vote for their own tool?

[Asmir @asmirkn]

@EXM7777 the mistake into a rule step sounds good but after a few weeks that rule file is gonna be huge and the agents will start ignoring half of it or contradicting it

[Asmir @asmirkn]

@fofrAI true, a file the agent overwrites every time gets ignored. make it an append-only progress log you can tail and you'll actually read it

[Asmir @asmirkn]

@amritwt yh great startup to me means the token bill is the one thing nobody wants to look at in the board meeting

[Asmir @asmirkn]

@gregisenberg how does the bot tell a stalled claude code agent from one that's just thinking through a long task?

[Asmir @asmirkn]

@OfficialLoganK @GoogleAIStudio When2Call is the tool-calling one, whether the agent should call a tool or ask, and TypeSafe Jev beats it there, 82.0 vs 79.9

[Asmir @asmirkn]

@nickvasilescu wait if hermes runs any model underneath, doesn't grok's voice change the second you swap the model? the personality is mostly the model itself https://t.co/F5uXUKLYp9

[Asmir @asmirkn]

@petergostev wait what is the +30.4% actually measured against? is the baseline fixed or does it change as new models join the leaderboard

[Asmir @asmirkn]

@hooeem an agent can watch 112 streams in an afternoon now. what you chose to look for in them is the only proof left