I swear everyone is high on their own supply
“Agents can’t write complex software in my codebase”
“I formally verify everything”
Ask them to show you the code
@simeonGriggs It’s insane the world we live in right now, where one day a company is top of the world and then a few months later, like, do they even still exist?
Literally no idea
Thank god MCP is dead
Just as useless of an idea as LLMs.txt was
It's all dumb abstractions that AI doesn't need because AI's are as smart as humans so they can just use what was already there which is APIs
slopped together a mac version of @zeeg 's cc-switcher for when you have your subscriptions purpose-isolated and you're not quite tokenmaxxing to need CLIProxy
just point your agent at https://t.co/YWJ0IqY8Au and tell it you want that
I think what you're just seeing is Opus 5.5 is really good, and the current generation of Frontier models are really good, and you can do a lot, you can get away with a lot. But if you look at anything that ships, there's so many loose ends because they're just not well defined and/or the scope is just really large
But it's really different than, like, oh, I've got a big production system with, you know, millions of users' data. I'm going to go make a change and trust that it's correct.
The way I describe this to everybody is you're optimizing for pre-work. You want to get as much done and as much automated and as much efficiency in your sort of human in the loop cycles that you can achieve more or you can achieve the same with less effort.
That's vastly different than you'll never need to look at code, though. There's probably some edge case scenarios where you could truly automate verification, but I think for the vast majority of software in the world, that just doesn't exist
I think the big thing is, though, like, am I going to go review the code of a front-end component? Probably not. Why would I? Is there value in doing that? I can probably analyze performance automatically, but I am going to review the visuals of that component or maybe a GIF slash video of the interactions or something
But that's very specific to, say, a front-end component. Like once you get into something like core infrastructure, like what are you going to do to develop trust that the change is okay? Well, yeah, I can make the review simpler by making the explanation simpler to follow, more efficient, but the code might actually matter
It's not a function of a token budget because at the end of the day, there are not predictable outcomes. If you described exactly what the change you wanted was fully to a spec, sure, you could verify that, but you don't. You usually type one sentence and expect the agent to infer either from existing context or from, frankly, mind reading.
So, unless we're willing to spend all of our time writing really concise specs instead of writing really concise code, or at least reviewing and iterating on code, I don't see anything possible that can change here.
@aidenybai imo lauren's thing around "develop trust with the agent" is true, but also i know that there are a large number of common scenarios where there is almost no level of trust the agent could create
@aidenybai i think its really about what the code is and where it is
changes to sentry? all human review (at some point we'll auto classify to reduce that)
changes to internal service with no customer data and no impact to uptime? depends on how annoying it wil be if its wrong
@Lermatroid i got it some years back and for the price it has been great. at the time tho there was basically no alternative to niche in that pricing range, and i think theres a few now
zero regrets for sure. i might buy a key when next iteration comes out, but thats real no need