I admit, I expected something different behind this header

smilarly great signals: - managing agents is like managing a team of humans - but i have review agents - spec is all you need

aha

Why are we calling coding agents "harness" now? Is there a semantic difference, or are the Fantasy nerds taking over the terminology?

Thanks - I hate it! android-developers.googleblo… via @google

the bastard maintainer from hell

it also ignores compounding effects on bugs, code quality, and maintainability (which matters to agents just as much, if not more). agents love nothing more than adding technical debt and unnecessary abstractions, resulting in bloat. bloat that agents can currently not handle.

if you're one of those breathless devfluencers perpetuating this garbage, fucking stop. it's a cute way to generate clicks to be converted to your crypto grift, but it ultimately sets the culture of our craft. like the rest of us, you have no fucking clue where this is going, but your clout gives you undeserved authority, not because you are producing anything of value, but because you scream the loudest. fucking stop.

it's funny, because yesterday, that same master clanker stomped all over the large clawdbot codebase, reverting fixes, breaking CI, because it can no longer grasp how stuff fits together. and nobody would have noticed. entirely fine in the clawdbot context, anything goes. but eventually, we need to stop kidding ourselves and realize, that the "engineering" part in SWE is still a thing, and the only thing that keeps our industry from imploding on itself when everyone and their mom gives in to the clankers 100%.

agent orchestration is a bit like hadoop hype at best you get some kind of parallelization but at what cost? still have to solve coordination do you know exactly what you want ahead of time can formulate it as directed acyclic graph? if you can solve that just prompt it

When Meta trains it models on 80 TB of pirated books from LibGen and other platforms, it's called 'fair use', without them having to pay penalties and / or receive some form of legal punishment, as proceedings are ongoing. When Aaron Swartz downloaded 70 GB of articles from JSTOR in 2010 he was facing a $1M fine and 35 years in prison, before taking his life in 2013.

Unpopular opinion: Unless you’re just prototyping, you should aim to understand as close to 100% of production code generated by LLMs. Yes, all of it. Effective mental models are still important for humans to sustainably maintain and evolve a codebase via prompting alone.

Use AI. Don't treat it like a god. Don't treat it like bs. You AI > AI. If that's not the case, invest more in your skills. Both your CS skills AND your AI usage skills

Man, this guy sucks and just can't shut up x.com/badlogicgames/status/1…