Why Bluefin is all in on agentic development | Bluefin

[giscus[bot]]

Why Bluefin is all in on agentic development | Bluefin

Is AI good or bad for open source? This is 58 minutes of me explaining what I've learned on this.

https://docs.projectbluefin.io/blog/bluefin-agentic-development/

[castrojo]

Huge shoutout @joshyorko for being the first person I remember to use an agent on Bluefin.

I remember thinking, wow. This changes everything. I have no idea what to do with this thing. And it was like, that weird awkward moment.

Soonish you'll just point an agent to an API endpoint and the thing will coordinate for you - now we can really figure out how to make this awesome!

[joshyorko]

@castrojo ty for the shout out. Been building along with you and learning from you and team. Seriously hive has changed the way I think about everything. Ty for bluefin, it is a daily party of my life now and I don't think I could have gotten where I was professionally just in the last year without the patterns this project has taught me. So ty seriously.

[castrojo]

The discourse around this is so frustrating. We have real data. You can see when James first connected hive to this part. The PR reviews on those models were so awful. Look at the latest reviews. You can literally see them getting better.

James and Andy worked together and now they're getting there. And the reviews are are at least as good at what we used to have. I speak into a microphone.

I say "Review mode" and my computer beckons my commands. And shit opens up on a terminal. And I feel like Geordi Laforge. I grew up with tube televisions. And now I am Geordi Laforge.

I'm not giving that up. I'm not ever giving that up.

[davestgermain]

If I look at the issues and here, it seems like even fewer people (or agents) are actually paying attention or interested in resolving serious bugs with the new image. And with the "benefit" of AI, fewer humans will understand how it all works...

It's still not possible to switch to the new bluefin-lts image because builds are broken. And if you do manage to switch to the image, it might not even boot.

I've used bluefin and bluefin-lts for a while now, but I'm beginning to wonder what the S in LTS stands for, especially if the system is one mis-prompt away from an AI agent messing something up and automatically pushing junk out (or removing functionality) to users.

[castrojo]

Is this a joke? The ARM builds were broken for months on the old images and none of the ARM images have every been GA. None of us have ARM gear.

You have the hardware and haven't send a pull request, and I'm spending money to fix YOUR problem. Find something else to use.

[davestgermain]

What are you talking about? The AMD64 build is failing, not the ARM build.

[castrojo]

You're blaming AI for no one caring about LTS. The use of AI doesn't change any of that.

[castrojo]

I apologize for my aggressive tone!

[castrojo]

Screenshot thing caught a bug, able to fix and push now! We'll have fresh Dakota builds tonight!

[castrojo]

Progress Report

People showed up for Dakota today! Onboarded a few folks and got a ton of feedback. This dramatically simplifies the structure. This is rolling out now but is not true yet. This will ultimately be in the contributor guide docs.

Let's start from the beginning. You want something in Bluefin, either a feature or file a bug.

• For technical things we'll push people to ujust report, which will be all whizbang (out of scope for this post)
• But say you want a new dinosaur or a new flatpak or something. You will file an bug in projectbluefin/common. You will then go in this order:

Flow

• 1-triage
• 2-discussing
• 3-human-queue
• 3-clanker-queue
• 4-review
• blocked
• hold

---

How this works.

1. 1-triage: We'll have you fill out a template here.
2. 2-discussing: People discuss and drive concensus.
3. A human being decides:
   • 3-human-queue for humans
   • 3-clanker-queue for the bots
   • The workflow will enforce this, it's either one or the other - this is currently Bluefin core contributors but we want to grow this circle so if you want to PM please lmk. It's just adding a few labels and being linux-clever. Ideally you're training each other up on this
4. 4-review: Standard PR review lifecycle. Requires +2 approvals from active maintainers before merging into the testing branch.

• At any time any person in the chain can call in clanker-queue, reviewers, etc.
• If a human wants to implement something using an agent, we don't care as long as the quality is good. The clanker-queue is for menial tasks that Hive will triage for us. It's designed to put the clankers to work.
• YOLO Mode (Opt-In Checkbox): Issue templates now feature an explicit checkbox:
  - [ ] Opt in to automated clanker review - this is basically "I give up maybe some kind soul will spend tokens on me."
  • I'm planning on measuring this as well. We'll see how many people actually care or don't care about AI when your wifi stops working. This will be a fun metric!

A few things:

• Strict Opt-In Enforcement: Hive has been updated to exclusively target 3-clanker-queue ONLY. Unprompted bot interactions and random background comments across issues and PRs have been completely removed. If a bot is annoying you tell me.
• All other labels are gone, we're simplifying the whole thing. 5 numbers and a few block labels.
• We're working on moving to this now so parts of the factory will be bursty.

[castrojo]

The hive part of the factory, aka put the spamming bots to work:

• it's not anonymous, it's basically good-first-issue for bots - the person's work is tied to their account
• let's everyone with tokens participate - we don't care where you get inference from. we'll do a "Donate to Bluefin with OpenRouter".
• the ujust stuff for contributing hive needs work, we need something like ujust command-clanker or something ominous and it should run the hive container via a quadlet all your system with full lifecycle management however you want
• Local models are VERY CAPABLE to contribute things. Hive coordinates, and we'll still use SOTA models to fix up the docs and continue to make more things deterministic.

Rate limits so the factory doesn't outpace contributors:

Full metrics, leaderboard, bling flexing your stats, etc.

[castrojo]

Also just realized this is just distributed.net and BOINC/SETI etc. but for clankers. Like I remember trying to put that on every machine I had influence over to help Team Ubuntu lol.

This is the same thing except it's fixing our computers. This is just the crude first step. And literally at a time when local models are getting very good and we're dealing with a well documented operating system. It's also important to remember that not using a thing like hive or swapping it out for something entirely different is the entire point.

Also this draws a clear line on what's what. I think it makes the value of the human gate go up tremendously. If you see an issue that's human-queue, it means it's something that demands the skill of a maintainer. So if your idea gets here, congratulations, you've made it past the drek of the maintenance around here, heh.

Also this method gets us both. First off it cuts out most of the spam off the bat. If you file an issue and you want human only, human-queue. We can now have that. That's a nice thing to have these things spam too much even when you tell them to shut up. And Andy added the lockin feature in hive itself. That sucked I can see why people hate this.

However, you might be new to linux and you want something that works, you can totally clanker-queue your issue, and some kind person donating some local machine time will take the data from your pastebin and run it through their local model with the expertise of the collective. And then post the findings back in a public issue. It's not even "telemetry" it's just automated copy and paste. And now we can put that noisy part to work because it doesn't take too many people in the swarm to pick up your issue. I'm willing to bet it'll be within minutes in some cases. The potential here is just too great to ignore!

Much of this work is just copying logs around it's not Opus. You need lots of local inference. 100k live bootc devices, most of which are Bazzite. That's a lot of GPUs man!

This is BOINC but you're working together to make Linux better. And there's a tmux window right in there for you to go in and see what's going on and figure out how to make it do your bidding. And all these skills are valuable in the marketplace.

[castrojo]

This also solves the "why do a gaming Bluefin in the first place?" - This is the same OGC kernel that Bazzite uses and designed to push things to the upstream Linux kernel.

[lore note]
In the lore the Bazzite kids will ride titanfall mechs and our people ride dinosaurs. But having a mech rider on a dinosaur makes sense in some kind of weird subclass kind of way. A healthy ecosystem would reflect some kind of reason to be allies, maybe the dinos are really good at sniffing out the powercore things the mechs need for power. Raptors are clever, clankers are not. It's a good combo, like having different classes in an RPG party.

It's more for the busy parent that games occassionaly, and now that it's literally just a checkbox? Easy win I think.

(Since we're doing art for different aspects of Bluefin when something is lore relevant I have to keep notes for the artists later. This is afterall a comic book story now too lol.)

[castrojo]

Update - here's what the review/contribution screen for people with agents looks like so far, you'll do a ujust review. This launches a premade VM/container to process the clanker queue based on permissions, roles, etc. Then hive will assign you work units:

We put goose in there and then hive handles the prompt, we add the right repo skills and agents so that it has everything it needs to do the scoped task that is has been given. The more context we add in here the better off we'll be. As people add their results we can keep documenting and making more parts deterministic for better results. Then we just keep looking.

The idea is to give people's agents the proper context - we hope to show that putting the machines to work with the proper tooling can lead to better results than spam, heh.

It'll passthrough your creds so you can use pi/opencode/whatever for this, but for now I'm working on a focused one here with goose/copilot.

[Danathar]

First thing - I'm assuming there will be some sort of video from the contributor's perspective at some point? I think that would be really helpful.

https://hosted-projectbluefin-knuckle-gjvq.hive.kubestellar.io/contribute

URL is awful but it's mostly working, point your agent there and hive will send you work. You can use whatever agent you want. It takes about 10 minutes and then when it's done it will either file an issue or send a PR. It's designed for you to lazily queue up things in the automation lane, so you shouldn't see some complex implementation or anything like that.

If something doesn't work tell it to report bugs and they'll get fixed. The onboarding page has a bunch of little bugs someone can just do an audit/gap analysis and report it to hive - agent issues are welcome and accepted.

Thank you!

[Danathar]

Well, got it to work with claude, but using goose with Qwen3-coder:30b...it just burped and couldn't handle it

[castrojo]

We still need to adjust timeouts and stuff. For locals I think I'll play with sending them to testsuite for a while.

We also can skew towards labels so that nvidia things end up on nvidia reviewer queues, etc.

[Danathar]

We still need to adjust timeouts and stuff. For locals I think I'll play with sending them to testsuite for a while.

We also can skew towards labels so that nvidia things end up on nvidia reviewer queues, etc.

Thanks. Some sort of routing feature could be useful too. The ability to have claude do the heavy thinking and then have it pass it off to a local model would be great.

[Edit: Ignore...evidently I CAN do this with goose and codex-acp]
The other thing, I'd LOVE to be able to use my ChatGPT plus account and use 5.6 Luna high or xhigh. I've kinda fallen in love with that model cause its decent and CHEAP. You can get a LOT of great usage out of luna. Of course, maybe Anthropic will come out with Haiku 5 someday....maybe....

[castrojo]

5.6-luna on high/max is my favorite as well but I am testing kimi:

https://github.com/projectbluefin/review has my container: ujust review-container kimi high

Then just tell it to fix whatever and then file the issues and we'll queue them up. We want something like ujust donate someday but for now we can work out of here.

[castrojo]

Man dude ....

[joshyorko]

@castrojo i love it i see mine in there

[castrojo]

Nice work! We got so much done this week! We're going to be set up nice for the weekend! If someone wants to rubber duck that /contribute page and file issues in kubestellar/hive they would appreciate it! There's other projects signing up: https://hive.kubestellar.io/

And it's all OSS (Hive is part of Kubestellar, a CNCF project) so you can run your own. I kind of wish I had like 6 of those AMD AI stack things.