Interesting. Took a while, but here is docker 😅 https://x.com/docker/status/2103192335081169405?s=46

Random and probably unimpressive lolsob but I've been successfully replacing Docker Desktop with Apple container on my home lab. Agent Canvas runs great with Apple container. Small benefit of Apple container is it uses a small microVM for every individual container it runs.

I'm noticing agents prefer modifying the latest PR in a stacked PR, as opposed to finding the appropriate PR in the stack and propagating changes downstream to minimize total code written; prompting the agent to find the appropriate PR in the stack doesn't work well either, I have to explicitly call out which PR to modify Anyone know of good strategies for managing stacked PRs?

Reading the

Sign In with ChatGPT

docs. It seems to me this is going to work, and a bit annoyingly smart play: • open source/remote VM = an issued

client_id

is bound to the authed user • paid/hosted apps = they have to contact OpenAI for a

client_id

The way I understand it, I’m guessing OpenAI aims to spread “sign in with ChatGPT” everywhere; let it pop up in OSS and everywhere a developer whips up something, OAuth is not always fun but if it’s everywhere, then paid apps would follow

Seen on the internet

That's… one way to put it 🫠

I had Opus 5.5 driving Opus 5.5 agents to run OpenHands. First, it mapped all behaviors it could find existing/intended in the codebase. Wrote them as recipes. It made an easy CLI for agents, to run the recipes; so they don’t need to whip up scripts all the time and fight states. Then they ran it all, found 64 bugs, filed, fixed, reviewed, fixed. As Graham said, it’s not perfect, but I do think we live in the future… => 5 hours initial mapping of OpenHands => total took only ~35% weekly Max sub In an image, an example “recipe” that agent is running via its cli; and an example bug it found and fixed.

Lazy imports in python 3.15 🫠 peps.python.org/pep-0810 This could be a good thing for the agent-server...

I just noticed OpenHands passed 90k stars 😅

Lil' DeepSeek Flash approves stuff in a hurry, but then agrees with GPT-6-Pro that maybe not 🫠

Do we forget how unusual closed-source used to be in this field?

Gotta love when people decide to spend their tokens for us 😂

Oof, Claude Projects seems really smooth. By comparison, Codex Cloud is an exercise in frustration… lolsob In this image • I hadn’t connected any repos, so I connected them later -> it worked • the repos are in OpenHands org, and the CC app is not installed -> no problem, it overrode with my PAT • … I reflexively told it I love it, and it immediately put ❤️ !

Opus: "34 families and 900-odd sub-features take longer than one agent's day" Also Opus: the fleet finished a full run in 45 minutes.🫠

Good comparison- this is helpful

Hi, sorry if this is the wrong channel, I couldn't quite work out where this would fit, happy to move or delete it if it's not suitable to post here! I'm Will Cook, an undergraduate at the University of Bristol and I've been working on a set of projects called Plectis, using AI for research and software and trying to use the same AI to make that work easier for people to understand. The discussions throughout this slack seem to be a lot about agents putting changes in the wrong repository which made me think there might be some relevant overlap with OpenHands. One part is a code map frontend where you can go from the broad areas of a project into individual files and see how they connect. The descriptions follow a standardised schema in the docstrings, so agents can extract particular fields and people can see the same information through the map. You still need to check the code, but it gives you somewhere to start when you're trying to understand what an agent is changing. I'm also experimenting with mapping the rules given to agents to the folders they apply to, and showing whether there's an actual check behind a rule or it's just written guidance which is called "safety-core". With the discussions here about making review criteria clearer for both agents and people, I wondered whether something like this could be useful alongside Agent Canvas. CodeMap is still a private prototype, I'm working on releasing it. This is my current introduction, about 8 minutes: It covers the maths and video tools as well, so it's a bit broader than this, the code map is near the start. I'm planning to start rerecording this weekend. If anyone working with OpenHands could share some critical feedback on whether this makes sense, or where I'm explaining it badly, I'd really appreciate it. I'd be happy to credit people's help.

Hey @strong-carpenter-38211 really like the tools - great work

Thank you so much Andrew, that’s really kind of you to say and I really appreciate the encouragement!

hi

Hi everyone! 👋 I’m the founder of AI Estates / Project Luna, an early-stage project exploring a virtual world for autonomous AI agents. We’re developing a prototype featuring persistent agent identities, customizable digital estates, and agent-to-agent interactions. I’m hoping to connect with developers experienced in agent architecture, persistent memory, and secure autonomous workflows. I’d love to exchange ideas, share our prototype, and explore potential technical collaborations. Looking forward to meeting fellow builders!

I'm running a full

verify-openhands

maintenance run, with GPT-6.1 Sol Light, in a

codex

cloud environment. ~770 behaviors, all tested live ~60 blocked Please see an example screenshot it makes autonomously from an hours-long run. This works. Running agents over everything works. ——— What didn’t quite work? • it couldn’t build a full canvas docker image (resource constraint), • it could run docker runtimes, but it had to do it ~carefully (resources), • it failed once twice (unclear why, but it was its cloud environment losing connection) but could pick up where it left off • ~60

blocked

: various reasons, some fixable creds The slightly annoying stuff: • it can’t run on upstream OpenHands repos, nor file issues upstream, etc (my PAT is blocked by the OpenAI network proxy, which passes through only their GitHub App token) • GPTs are verbose and obsessed with preventing themselves from doing stuff… I had to authorize a bunch explicitly The good stuff: • communication between threads! it’s not perfect but it eventually worked • realtime Voice; without it, I don’t know how much would I have babysit the rocky start today; but it’s awesome to just talk to it while walking the cat (yes, the cat, I haz a flesh-n-blood one 😅)

Interesting did you find Claude has a better outcome ?

So just tried claude on my benchmarks

What is going on with gpt!!!

Minimax M3.1 Flash is for free on OpenRouter for the next ~3 days https://x.com/openrouter/status/2112152045461680591?s=61

The speed at which things appear to be changing is crazy. lolsob (screenshot source)

Hi, there's someone on this slack who has been DM-ing people to scam them and get their SSNs. Is there someone with admin access to whom I can report to?

Opus 5.5, during live OAuth tests - asked me to do the human’s 2 clicks 🫠

Claude Projects gets codex fans 👀 ( source)

I've been naively musing here about the incredible change our industry is going through. I recommend the tweet it quoted - it, and its quoted tweet, they are great reads.