I was all in on sandboxes and safehouse for my agents but the moment I got into iOS development it felt like my hand was forced to just run Claude / codex / pi directly on my machine because nothing else could do the dev loop.
It’s been a painful reality for me, I’m going against core pieces of how I feel I should be interacting with agent harnesses and yet, I need to get the work done so
Hey, I work at Docker and my team works on mcp integration with sbx. A solution I've been trying is this:
1) Enable the xcode mcp server: https://developer.apple.com/documentation/xcode/giving-exter...
2) Add the xcode mcp server to sbx: `sbx mcp add xcode --command xcrun --args mcpbridge`
3) When you create the sandbox, use `--static-mcp xcode`. For example: `sbx create --static-mcp xcode claude .`
Make sure you have at least v0.38.0 of sbx. This makes a bridge from inside the sandbox to the xcode tools on your host, so be aware that it can run whatever tools you give it on the host. But the agent itself is still sandboxed.
Also had this pain point as an sbx user. Given the risk this adds to the host, would be great if there were more docs on how to setup kits to make it safer (e.g. disable yolo mode).
I am super curious how much in your opinion the extra prompting is useful?
>often harnesses also mention in their system prompt locations of markdown files that the model can consult if the model thinks they might help
>that hint alone as part of the system prompt can be strong enough to make the model read in more tokens than would have been necessary
I purposefully do this as I imagine it is useful. In my project I am seeing solid adherence to norms and a deep capacity to iterate on completed features. Essentially for each feature I have the model make a {featureName}.AGENTS.md at the root folder of the feature, where it maintains what is going on.
I am moving between Claude Code and Codex atm, but I began this pattern when pi + kimi 2.6 was my main driver.
I do think prompting and reference files (e.g. for architecture, tech stack, …) can be extremely helpful. I also do this in my projects (challenge is keeping drift of these documents at bay).
What I wanted to emphasize is that whatever is in the context (whether system prompt or user message does 'steer' the model in a strong way, so everything in the context affects overall performance in a way. Even if it is 'just' net neutral it takes space up in the context window.
The context window is very very precious, everything that goes into it should help (not just hopefully help).
The challenge is coming up with good stuff to put into that context. A good agents.md file will be better context than whatever the popular harnesses have in their system prompt.
Also good to keep in mind that newer models are very good and more agentic than older models so they are better at exploring their environment based on the tasks you give them.
Thank you for this project - I am excited to check it out!
My current agentic workflow is having a github repo that is a workspace and essentially a single obsidian vault over it called agents. I modify with Claude code (or other harness), check diffs in...VSCode and read it in...Obsidian.
It's separate from my actual personal Obsidian vault (that I don't send to any AI providers), and is only for agents. It's been really nice on all sorts of varied projects and for performing web research. I also have a bun monorepo setup in the root with varied tools for search, fetching individual websites, setting up folder structures - etc.
But essentially, in my experience, the moment you tell the model that it's in an Obsidian Vault - magic happens.
I am so curious how this will play in with it all and am hoping this will improve my workflow!
Gah - this is unfortunate. This will have me quit as well - I wonder when they will begin to enforce this on the Claude Code front. I remember when OpenAI tried (maybe they still are?) putting this in for GPT 5 and this had me switch from them to Anthropic.
I'm curious what part of the charts would data centers fall under?
I'm assuming it would be under the Energy emissions category, but I didn't find anything particular around data centers or "technology" or "internet" or something like that.
Would it fall under "Electricity and heat"? Or just general "Buildings"? or "Commercial buildings"? Or am I way off base?
> For example if AI generates 2x of a utility function that does the same thing, yes that is not an ideal, but is also fairly minor in terms of tech debt. I think as long as all behaviors introduced by new code are comprehensively tested, it becomes less significant that there can be some level of code duplication.
We still run into the same issues that this brings about in the first place, AI or no AI. When requirements change will it update both functions? If it is rewriting them because it didn't see it existed in the first place, probably not. And there will likely be slight variations in function / components names, so it wouldn't be a clean grep to make the changes.
It may not impact velocity or stability in the exact moment, but in 6 months or a year - it likely will, the classic trope of tech debt.
I have no solution for this, it's definitely a tricky balance and one that we've been struggling with human written code since the dawn.
reply