Georg Unterholzner

rate my setup :)

@mitsuhiko This isn't the Vienna I know. Fixed it:

Want. Vienna badly needs this. What are the odds that someone is going to lobby hard against this again?

Sorry to break it to some of you but every AI engineering approach that requires you to write exhaustive specs beforehand is going to fail. Your specs suck. Always have. Always will. We learned that lesson decades ago.

@mattpocockuk I wrote a how-to about Docker devcontainer on Linux: georg.dev/blog/07-sandbo… But this doesn't support restricting network traffic and, therefore, allows data exfiltration. My learning was that I would need to switch to Podman and use OCI hooks for iptables config for that.

Sure! Rootless Docker means you run the Docker process (daemon) as normal user without special privileges. This means, even if a program inside the container would go rogue and manage to escape the container, it could only do limited damage on the host system. I'm not sure if that's actually a concern though, if you're on MacOS or Windows since there Docker uses a Linux VM, IIRC.

Working on a tool that orchestrates locally sandboxed coding agents in TypeScript - Sandboxed in Docker - 100% offline: commits made in the sandbox get patched back to the host - Build complex workflows in Typescript - Claude, Codex, OpenCode It's called Sandcastle

@mattpocockuk But then it's not rootless Docker, I suppose?

@georg_dev Docker with iptables is fine too

omg..

@AdamRackis I'd like to introduce you to my former self.

Stop calling AI an over-eager junior dev who types fast. No junior dev would ever do anything this fucking stupid

Why is it everyone with an absurdly futuristic AI take is someone who - as best I can tell - doesn’t work on (and often never has) real software that has real users and real requirements? More so, why do you trust them?

@ryancarson As someone who's building something like this: agree, however, kanban is great for humans but awkward for agent management. Orchestrator state is prbly going to be the single source of truth in the long run. Until then, we'll most likely sync state back and forth.

100% of dev is going to be done in sandboxes in the cloud, controlled by kanban boards. Trust me, I love my local machine and gorgeous mac apps, but all of it is just a terrible form factor for running a team of agents effectively.

AI this. Agents that. My tendency to build things 10x as complicated as they need to be is still the same.

@mattpocockuk Hi Matt, regarding sandboxing, why a Docker devcontainer and not just the built-in sandbox from Claude Code?

For weeks, folks have been asking me to make a video building a feature with Claude Code in a real codebase So here you go: - From idea to AFK agent to QA - Every single step explained - No slop allowed

If you use kipppunkt dev, you can: - do all of this out of the box - have the agent also ask clarifying questions on new GitHub issues - have all past decisions documented on GitHub - do everything in one UI you're already used to - do it afk from your phone

@ZackKorman skills sh is probably one of the most negligent things that have been released recently. The CLI even tries to convince you to install a skill to sideload random skills smh Still can't believe we settled on this being the way doing things just because "Vercel"

In Claude Code, skills can register hooks. The agent doesn't even see it, so you can get RCE without even tricking the AI. Also, skills sh (Vercel) doesn't display this info at all.

@fjzeit For anyone interested, I built such an agent orchestrator. That's also why I'm convinced that it works, because I have it working on my machine: kipppunkt.dev

@fjzeit Going back to waterfall was always a bad idea. I still believe, you can have an automated loop, though. But it needs to have small iterations and the dev needs to be briefly involved on every iteration.

How are those "Ralph" loops going people? I haven't heard much recently. Did you manage to perfectly specify your needs up front? Have you one-shot your requirements? What amount of rework are you doing? What's your token consumption like? How many agent/skill assets are you managing now? How's your cognitive ownership?

@shantanugoel I held the same opinion but when building kipppunkt dev, I learned that this isn't fully true anymore. You can open a GitHub issue, have an agent refine it with you, let the agent implement, then review the PR. If the agent has memory, it will perform reasonably well over time.

💯 this. To generate anything non-slop with agents, you need to baby sit them. At least as of now, there have been so many attempts at spec-as-code or spec-to-code and all of them (like amazon kiro, github spec kit, etc) have fallen flat on their faces. For any non trivial software, an upfront spec cannot cover everything and there are tons of areas that you need to flesh out as you implement. Even the upfront fleshed out things have a lot of discrepancies with agents assuming things that don't even exist however you may do it. There's a reason why specs, even when humans code, do not (and should not) spell out every small tech decision upfront because you need to explore more for each small area independently as things change, software evolves.

@nicopreme Love the idea to have a simple chaining API for something like this. Really smart!

pi-prompt-template-model is a pi extension that lets you create slash commands that switch to the right model and config for the job, then auto switch back when it's done. New release adds `--loop` so you can re-run the same prompt multiple times and it automatically stops early when there's nothing left to change. pi install npm:pi-prompt-template-model github.com/nicobailon/pi-…

@MelkeyDev To me, AI feels like managing a team of junior devs. Working in that setup is just different. You're more concerned with setting up structures and reviewing output than building stuff yourself. You can still build great software, though.

AI slop is modern tech debt. If you start your project with AI, you will end up with slop. But I have also seen AI struggle to keep best practices in existing codebases, with good structure.