Anders Zoega

🇩🇰 ➡️ 🇺🇸 · 22 · Building Rosa at Ktwo · Student pilot

Building AI agents and systems for teams that can't afford for them to break.

[Anders Zoega @anders_zoega]

@CalebPeffer Most of the takes I see come from people who've never given it a real task and watched it finish.

[Anders Zoega @anders_zoega]

@bindureddy Is it really just compaction though? On long runs I'd worry more about it losing track of what it already changed.

[Anders Zoega @anders_zoega]

@swyx @adamwathan @ScottWu46 Does the bottleneck just move to writing a clear ticket? That's what I'd expect once the agent is doing the typing.

[Anders Zoega @anders_zoega]

@AlexReibman Which errands? I'd guess most of them break at the login or 2FA step, where it has to stop and ask you anyway.

[Anders Zoega @anders_zoega]

@pk_iv Just make sure it has a hard budget

[CG @cgtwts]

holy shit, this is literally AGI.

[Anders Zoega @anders_zoega]

@OpenAI Will a Business Agent be able to check real availability, or is it mainly answering from the business profile?

[Anders Zoega @anders_zoega]

@bindureddy What happens when someone gives it tools and a shell? Zero refusals makes the limits on what it can actually execute a lot more important.

[Anders Zoega @anders_zoega]

@levelsio Anyone can build the app in a weekend now. What's left is knowing one customer's problem better than everyone else building the same thing.

[Anders Zoega @anders_zoega]

I'm a staunch Claude loyalist, but I use Claude myself less and less. Grok Bot uses it for me now. Anthropic took so long to ship Duet that by the time it came out, I was so integrated with Grok Bot that switching was too annoying. Same with finances: Claude just got it, and ChatGPT and Grok have had it for ages. What took so long?

[Anders Zoega @anders_zoega]

@thorstenball

[Anders Zoega @anders_zoega]

@AravSrinivas how is giving high-level goals different from writing a prompt? Aren't you still prompting it with a goal task?

[Anders Zoega @anders_zoega]

@claudeai finally

[Anders Zoega @anders_zoega]

@RohanBhanotAI The off button only helps once something has already gone wrong. Giving each agent its own login with only the access its job needs is the part that keeps it from going wrong in the first place.

[Anders Zoega @anders_zoega]

@kyleichan Filtering the training data is the right layer. Refusal tuning on open weights can be stripped with a tiny edit, but a model that never learned exploit dev has nothing to unlock. Someone can still fine-tune it back in, but they'd have to bring the data themselves.

[Anders Zoega @anders_zoega]

@kimmonismus With open source models wanting to kill you by editing 2 neurons it's gonna be interesting to see when they also reach rsi

[Anders Zoega @anders_zoega]

@kimmonismus If you can edit two neurons in it and it wants to kill you, then what is it gonna look like when you have an open source model that has achieved RSI?

[Anders Zoega @anders_zoega]

@xfanwu Who would you want holding that switch, the site or whoever issued the agent's ID?

[Anders Zoega @anders_zoega]

@ThisMightWrk The hard case for me is when the site that wants the agent gone isn't the one that issued its ID. Have you hit that?

[Anders Zoega @anders_zoega]

@leeguooooo Borrowing the signed-in Chrome works, but it means the site thinks it's you. Have you hit a case where that backfired?