This is because for some reason all agentic systems think that slapping cron on it is enough, but that completely ignores decades of knowledge about prospective memory. Take a look at https://theredbeard.io/blog/the-missing-memory-type/ for a write-up on exactly that.
It’s a self fulfilling prophecy. They’re extremely expensive so they must be good so they must be worth it. And because at that level measurement is extremely subjective it’s mainly about the vibes.
A vibe? It’s completely obvious AI slop with no attempt to make it legible. They didn’t even prompt out the emdashes. For such a cool finding this is extremely disappointing.
It's a fair question. I've had problems with Gemini 3 due to rate limiting, and I've been working on this for a while now. I'm planning Gemini 3 for a follow up.
It’s not groundbreaking in a technological sense. The codebase is actually a bit of a monstrosity. But it removed guardrails that were artificially put on these LLMs which suddenly gave it an entire new dimension and the timing was right.
I built this because I was curious what Claude sends to the API, how subagents get work delegated and what contexts look like. Interesting to see how small part of the context the user interaction really is typically.
I built this because I was curious what Claude sends to the API, how subagents get work delegated and how contexts look like. Interesting to see how small part of the context the user interaction really is typically.
Skipping the investigation phase to jump straight to solutions has killed projects for decades. Requirements docs nobody reads, analysis nobody does, straight to coding because that feels like progress. AI makes this pattern incredibly attractive: you get something that looks like a solution in seconds. Why spend hours understanding the problem when you can have code right now?
The article's point about AI code being "someone else's code" hits different when you realize neither of you built the context. I've been measuring what actually happens inside AI coding sessions; over 60% of what the model sees is file contents and command output, stuff you never look at. Nobody did the work of understanding by building / designing it. You're reviewing code that nobody understood while writing it, and the model is doing the same.
This is why the evaluation problem is so problematic. You skipped building context to save time, but now you need that context to know if the output is any good. The investigation you didn't do upfront is exactly what you need to review the AI's work.
OSS was already brutal for new contributors before AI. You'd spend hours on a good-faith PR and get ignored for months, or get torn apart in review because you didn't know the unwritten conventions. The signal-to-noise ratio sucked but at least maintainers would eventually look at your stuff.
Now with AI-generated spam everywhere, maintainers have even more reason to be suspicious of unknown names. Vouch solves their problem, but think about what it means for someone trying to break in. You need someone to vouch for you before you can contribute, but how do you get someone to vouch for you if you can't contribute?
I get why maintainers need this. But we're formalizing a system that makes OSS even more of an insider's club. The cold start problem doesn't really get any warmer like this.