Michael Collier

@holaxis

[Michael Collier @mpc2055]

@bcherny @delali @bcherny nested subagents have been an issue for me. I had >50 subagents running concurrently for a simple task because subagents kept spawning more layers. And it makes it worse that ootb, each agent writes to the other agent’s context window,so it’s a game of telephone.

[Michael Collier @mpc2055]

@jasonzhou1993 Yep. No wonder they are having capacity issues more frequently

[Michael Collier @mpc2055]

@burkov Whenever I explicitly want to search a lot of sources (including academic papers), I reach for Perplexity. I have both chatGPT Pro and Claude Max, and still use Perplexity at least a few times per week.

[Michael Collier @mpc2055]

@burkov @tvbweb2 This is a big stretch. Your microwave can’t currently have a “conversation” with you about its potential consciousness.

[Michael Collier @mpc2055]

@nummanali I’ve been begging for an increase to the 6. Help make it happen @nummanali !

[Michael Collier @mpc2055]

@startupideaspod I’m a huge fan of Opus, but most of what is written here is just plain false.

[Michael Collier @mpc2055]

@burkov Not sure if you realize what a gift a $200 per month plan on either Claude Code or Codex is. Thousands of dollars of inference, if you are skilled at using it. If you are working on economically viable things, $2,400 per year is peanuts relative to the potential value creation.

[Michael Collier @mpc2055]

@gkcs_ Wasn’t AI-2027 written in Apr 2025?

[Michael Collier @mpc2055]

@TheWorthyHouse @JoshDaws This a dangerous position to take. Absence of evidence is not evidence of absence. Absolutely nothing he said is inconsistent with my daily personal experience.

[Michael Collier @mpc2055]

@mattshumer_ On point. You described my every day experience. Thanks for writing this up.

[Michael Collier @mpc2055]

@thsottiaux @rafaelobitten I’d be ok with a $1k per month plan to not have to worry about usage limits.

[Michael Collier @mpc2055]

@thsottiaux I would very much appreciate living the 6 concurrent subagent limit. This is the single biggest reason I default to Claude Code. If you changed this, I would be Codex first. Pretty please?

[Michael Collier @mpc2055]

@thsottiaux @thsottiaux “staggering scale” with a 6 concurrent sub-agent limit doesn’t line up. Please remove this limit so we can actually supervise agents at scale. This is the only reason I still reach for Claude Code.

[Michael Collier @mpc2055]

@steipete Intent-Driven Development

[Michael Collier @mpc2055]

Opus 4.6 is available for me!

[Michael Collier @mpc2055]

@VibeCoderMcSwag @burkov Dude I swear, the shit he says gets dumber and dumber. It’s amazing to me that someone can write a book on LLMs but have absolutely NO CLUE when it comes to applied AI.

[Michael Collier @mpc2055]

@burkov Yeah, this is EXACTLY how real software dev works. Write a spec and hand it to an engineer, then evaluate the outcome after the first attempt. Why hasn’t anyone thought of this before? If one can write a spec and its perfect in one shot, they MUST be a rockstar vibe coder 🤣

[Michael Collier @mpc2055]

@infoxiao Clarity is kind.

[Michael Collier @mpc2055]

@vasuman Every day, with prompting alone, I spin up 10-15 agents in parallel to do extremely complex tasks. Sub agents are useful, but I prefer to have the main agent (orchestrator) create agents ephemerally by prompting them itself in order to meet my goals.

[Michael Collier @mpc2055]

@vasuman @vasuman this post is now over a month old, so ancient. But if you plug the prompt into Claude Code for your specific use case, you should be able to see the magic happen. Since this post, Claude code has enabled background - ask for “foreground agents.”

https://t.co/XOAKQqfzXI