nathanqueme

Hey HN, I'm Nathan 22.

In the last 2 years, I've built agents in production that worked inside people's daily apps like Gmail or Google Sheets.

I've always wanted to have full control and ownership over my agents' tools and that led me to create an open-source lib in python that lets you do that.

There are many ways agents can go wrong in production, and it's not always clear what they can do. So Charter lets you answer that, to understand your agent's scopes and enforce rules.

You can get started using examples in the repo, simply by running a python script, and it's free.

Hope you find that lib valuable, and I'm curious to know how you audit your agent tools today.

When you say it costs 57% less to use Halv do you mean this is due to routing to a cheaper model or thanks to compression?

I built a library to have a way to check Claude Code's work against API docs, and asked it to write many tests on top, but even then I found Claude Code would paraphrase and drift from the docs. I also found that it passed its own smoke tests but failed when testing against the real APIs, so reviewing Claude Code's code is tricky.

I'm junior engineer myself, 22, just graduated from a bachelor degree in AI, and I think that one junior can beat a senior with a Max plan.

I think it all comes down to what the they did during their studies and how they've spent their time. So I'd judge them based on the things they've built and their experience using AI tools. These days you literally have top US companies like Sierra asking their candidates to vibe coding a project as the coding interview.

I personally spent a lot of time and money building with Claude Code and Cursor. Probably more than you would at most companies, especially an intern: full days (weekends included) and about $5,000 in 2 years on tokens. During my sophomore year, I went so hard building agents with AI, I skipped all my classes and barely passed my subject, which got me Tier-1 VC inbounds. So the junior engineers in their early 20s can be more proficient than senior devs.

A terminal also tells the model when it failed. Exit code, stderr, a stack trace. it reads that and tries again.

The apps normal people use mostly don't. I spent two years building agents that worked inside people's gmail and calendar, and the failures that hurt all came back as 200 OK. - Slack puts {"ok": false} inside a 200. - Google ignores query params it doesn't recognise, so i sent max_results instead of maxResults for weeks and every filter was silently dropped.

The agent says "done", the user finds out later, and they stop trusting it. So i think part of the gap is that agents look a lot dumber when nothing tells them they got it wrong.

AGI is like water, and these surfaces are the plumbing. Agents and the internet's existing APIs are how AGI will actually be delivered in the short term.

But you're completely right: our current APIs were optimized for humans, and some like Excel date from the 80s. So in the long term, agents will highly likely create their own interfaces optimized for when no human is in the loop.

Which is exactly why companies stick with current APIs for now. Fable 5.1 already emits alien-like reasoning traces, and the recent OpenAI agent swarm accident just proved this.