Hummingbird Community Meeting Notes - 25 June 2026

Project Discussion
1

Hi folks! :wave: New community architect for the Hummingbird family of projects here, and I’ve got the notes from yesterday’s community meeting that @fatherlinux ran. We had a bit of a hiccup again with the recording, but we hope to have that up somewhere soon (I’ll add it to the thread here when we have a link for you).

Full disclosure: I’m putting AI-generated summary notes in here to insure everything at least has a chance of getting captured should I have missed something, and then I’m editing and adding my own notes.

Topic 1: Growing the community

  • Role Introduction: The meeting introduced the new community architect, who is leading efforts to build a community vision for Fedora Hummingbird Linux and Project Hummingbird (hardened images). (Hey, that’s me! Feel free to reach out with questions.)
  • Goals: The primary goal is to foster community participation in documentation, development, and image requests.
  • Transparency and Content: Efforts are underway to improve the transparency of the project, including creating blog articles and videos to help users understand the overall landscape.
  • Communication Channels: Fedora Discussion posts are the primary focus for community interaction to facilitate long-form discussions, with limited activity on Matrix channels and no dedicated Matrix channel yet. A landing page has been established at https://forge.fedoraproject.org/hummingbird.

Topic 2: Fedora/DevConfCZ recap

  • DevConf and Flock: Team members attended DevConf and Flock, noting significant community interest in the project.
  • AI Integration: Discussions highlighted constructive skepticism regarding AI, leading to a consensus on maintaining a “human-in-the-loop” approach for responsible AI usage.
  • Documentation and Pipelines: The current documentation and Git repository structures are enabling faster development workflows, which were described as more efficient than relying solely on AI agents.
  • “Chunka”: This tool, used for container image construction, intelligently “chunks” images into layers (e.g., one layer per RPM). This improves network efficiency and deduplication on disk, as redundant data is not re-downloaded.

Topic 3: Image Requests feature

  • Image Requests: A new preview feature on images.redhat.com allows users to see, request, and track the status of images. (Note: Yes, this is still on redhat.com, but that’s only because we’re still trying to figure out where it lives. You’ll find in the recording the more in-depth answer on this question, but suffice to say it’s mainly because we’re still trying to get settled and figure out how to move everything over to a community site.)

Topic 4: Shipping updates

  • Recent Shipments: Recent additions to the images include kubectl 1.36, OpenTelemetry collector, Tofu, Sift, and Bind.

Topic 5: LTS Strategy

  • LTS Considerations: The team is evaluating how to manage images for software components that reach end-of-life.
  • Rolling vs. Stable: The current model involves rolling dependency trees to minimize CVEs while maintaining API compatibility. A debate exists regarding whether to continue this rolling approach for end-of-life components or shift to a more stable model for edge-case users.

(Note: The recording accidentally stopped here, so what follows from here on out are my notes.)

In the questions that were coming up, it was clear we needed to take a step back on this specific topic and address the overall vagueness around the family of projects. What’s part of Fedora, what’s not part of Fedora, how will all of this change over time, how this relates to Fedora Atomic Desktops, and just what falls under the concept of “Hummingbird” here.

In short, we’re still trying to figure all of this out. However, Fedora Hummingbird Linux can be thought of currently as a host for Project Hummingbird containers; it’s different from, say, CoreOS in how it’s built (the factory; more on that in a moment) and the focus on mitigating CVEs within 24 hours. It may someday be a desktop, as well, or it may change into something else. Project Hummingbird, the container images to use to build containerized applications with, will hopefully move into Fedora along with Fedora Hummingbird Linux soon. That’s why Project Hummingbird is being brought up here and discussed so heavily.

Scott shared a graphic with the Hummingbird factory as an icon at the top, with two arrows pointing down, one to Fedora Hummingbird Linux and one to Project Hummingbird. These two projects are related by that factory. The factory itself is an agentic AI factory, but it’s not generating a vibe-coded operating system or vibe-coded containers. Here’s roughly how it works (I’m still getting all the details myself, but this is my understanding): When a human opens an issue to report a bug or request a feature, agents work on the various parts of handling that bug or feature request from (1) validating the report by choosing which scripts to run to spin up test environments, for example, to (2) ensuring that all of the relevant data is present in the report and asking for additional information if needed to (3) exploring possible fixes or paths to adding the feature. Then, agents (4) create patches that, finally (5) humans go and review and validate.

Right now, the infrastructure lives on Red Hat systems simply because it’s going to take a lot of time and discussion to figure out how to migrate it where it needs to live. We’re working on building up contribution pathways so we can get people access through a contributor ladder, just like everywhere else. More to come!

Topic 6: Blogging and project content

There was a discussion about where we want to have all of the project content. Docs obviously will make their way into Fedora’s docs as the project moves over and starts creating docs (or moves existing docs over), but we were thinking more along the lines of where we wanted to have blogs since, currently, there’s two projects with two homes. We want to encourage people to post things here to Discussions, but we also need more of a curated feed where we can ensure the technical details are correct before stamping them as the project’s blog posts. After some discussion, we decided that we’ll focus on Fedora Magazine, and we’ll also repost in other places like the hummingbird-project.io site using the canonical link tag (<link rel="canonical" href="your-link-here" />) to keep the SEO with Fedora Magazine.

We’d love your posts to end up on these sites! As we start to find more and more posts here in Discussions with the hummingbird or hummingbird-sig tags, we’ll ask you if we can help you get published. More to come!

Topic 7: Administrivia - upcoming virtual face-to-face

Scott wanted to let people know that there’s an upcoming engineering face-to-face for Hummingbird, so if you have anything we need to talk about, let us know!

Topic 8: Open Floor/Q&A

We opened the floor after this to Q&A, and we talked more about the great questions that came up about the projects and their relationships and use cases all during the LTS strategy discussion (notes from me about that a bit further up this post). I think I covered pretty much everything we talked about in that section of this post, but please feel free to ask more questions and correct me if you were there!

I think that’s everything.

Our next community meeting will be 23 July 2026 at 7a CDT (12:00:00 UTC) in our Jitsi meeting room (links and details are on our Forge homepage). I’ll get it up on Fedocal soon!

1 Like
2

Notes for the Week

Laura and I had a few discussions on Agentic Operating Systems.

What is an Agentic OS

I’m going to ponytail here - I read these articles and I think “I can’t think of a more convoluted way for people to explain Linux.” There are so many articles out there and most are bad. My short term goal with the messaging is in the spirit of a register article: “Linux community points out to techbros that the Agentic OS is just linux.”

Tenets of the Agentic Operating System

I started to write the big do’s and don’ts on a whiteboard, no order, discussion topics. Since so many people are just making things up we can just grab our terminology back and define them the correct way. :slight_smile:

  • Agents assist, and are never a hard dependency
    • Factory ops are well defined IaC - this is an orchestration problem not an AI problem.
    • The AI people selling “Agentic OS” - The idea of agents making autonomous decisions at this time is not an option. They are totally capable but why would you pay for that when you can have local models just flip deterministic blobs. And even if I had unlimited tokes, at this scale? It’s too slow and expensive I think the tokenomics will kill this.
  • Entirely gitops driven for operation, qa driven for design
  • Be totally self sustainable, even offline: build it’s own OS and service it’s own ecosystem - full supply chain security is a requirement not an option.
  • Recognize that an “Agentic OS is a solved problem technology wise. I am not using anything new.” This is all well documented Linux and ops. Everyone thinks we need to have agents solve cold fusion it’s just argoctl lol.
  • Self Improvement - the lab MUST self maintain and improve ← probably one of the most important.

The Lab

My goal is to prove that the agentic os is linux, k8s, and some yaml files - because even if we fail whatever wins will need at least those three things. I’ve got a dedicated lab for this and will post updates. Right now it’s 2 Framework desktops but adding a third - the usb4 stuff landing in the kernel is going to be really awesome for this. I’ll focus on the software:

I am running two nodes, on Fedora, one Flatcar. This lab will be a hodgepodge so we can find as many wins for as many people as possible. Mine will focus on CNCF-first tech.

  • Most of this stuff needs new kernels. Fedora is perfect for this, not only is the kernel up to date it ships the right usb4/modern stuff turned on.
  • Flatcar ships vanilla LTS kernels. Since this is a self healing cluster I spoke “Ensure the flatcar nodes are always on the latest kernels, check nightly” into a mic and now I have a fully deterministic upstream kernel workflow. And it even set up A/B testing on the cluster and will automatically stage, rollback, etc. This took about a night’s work.

Why gitops? I needed a way to get data “in” and “out” of the factory. No ssh, no fancy access rules, etc.

Here’s where the investment in bootc pays off. We’re just testing container lol. Every 10 minutes the lab checks for new images from fedora, bluefin, bazzite, aurora, ublue, etc. then it runs a battery of tests on all of them and then posts a screenshot, etc.

Christian Schaller pointed out the red hat framework for desktop testing and was thinking of ways we could use it for something useful for the community. So we took it all and started making a GNOME 50 suite, apps, etc. everything. We can do apps too, I think we’ll be able to test every aspect of Bazaar soon, etc.

And the last aspect is the self improvement. Each of the labs docs, workflows are either on their way to becoming deterministic or improving an existing skill. When we find something broken we use a frontier model and fix the design:

  1. Design: ~2 months ago. Basic seed: “ublue as first gen bootc factory, soft-fork projectbluefin to create an automated agentic factory based on … (5 years of data) and this red hat blog post, ignore the internet” - Opus 4.7, about three hours.
  2. “Day to day” operation of projectbluefin images, centos/fedora/gnomeOS - Sonnet 4.6, bulk of the time. Using it to review issues/PRs, analyze reports sent by contributors!!!
  3. 2 weeks ago - successful automated releases of all the images with Sonnet 4.6
  4. Wednesday this week - successful bump down to Google Flash models, saves about 40% cost, results are fantastic due to factory doc self improvements. (Each contributor uses an end session skill with instructions to capture and self improve the factory)
  5. Friday - successful bump down to GPT 5.3 - codex for factory ops.

I am 100% convinced that we’ll be driving bluefin with local models sooner rather than later.

What sucks about it?

(I am here to prove it’s doable if you are an expert on this tell me what to do lol):

  • I am not a QA guy that testsuite is a result of AI fever dreams and asking people who know. It’s real though, look at all that red, I can confirm it’s real!
  • I am also not a local model training guy - we need to find an AI model nerd that will take the leading local models and tweak them for this. Everything on the system is deterministic and well document - with the right set of experts we could make much faster progress here.
  • I am likely doing kubevirt wrong, I don’t know the happy path, it works but I’m still unsure I think KubeVirt is a critical play here. Doing all the magical hardware passthrough stuff is useful for desktop AND AI workloads because those are going into VMs.
4 Likes
3

I agree with your assertion that it’s all just Linux. Always is, always has been. I recently went down the rabbit hole of trying to think of the Agentic OS hype cycle and that lead me to the question: “how could we bake OpenClaw into a spin of Fedora Hummingbird?”

After arguing with my coding agent (goose) with some combination of Gemini models, OpenAI models, and Anthropic models for a day or two about this idea, “we” came up with the general idea of exposing tool calls to OpenClaw via “portals” that are baked by native system calls/commands or over dbus with various policy gates to ensure the agent is only allowed to do certain things as per PolicyKit and OPA. Also all agent execution is sandboxed via OpenShell.

The thing I’m still debating in my own mind is basically: “is this a good idea, is any of this useful, does this make sense, and is this even necessary?”

I’d love thoughts/feedback.

1 Like
4

The other thing that you mentioned is the testing via AI. If you’re down to the “lesser” frontier models, you’re likely at a place where the Gemma4 31B model would do what you need locally and to the best of my knowledge that will run on a Strix Halo system pretty well. I’d be very interested in results there. Unfortunately I lack the hardware myself but my current working theory is that the latest Gemma4 and Qwen 3.6 “small” models that can reasonably be run on local hardware are probably good enough for what you need that you don’t need any tuning/tweaking. Iirc both Gemma4 and Qwen3.6 are scoring better than Sonnet 4.5 on almost every benchmark and are within a tight margin with Sonnet 4.6 in certain benchmarks. I say give it a try running them locally with Ramalama and see how it goes :slight_smile:

2 Likes
5

You’d be correct! I got three of them haha. Once the images are done this is my next step,. gemma4 is the one we’ll shoot for first.

I think that this is a clear evolution, you’re not crazy!

separates “run code” from “change the computer.”

This should be a tenet somewhere. I’ve wanted this for a long time and have been shopping around the idea of an OS centric MCP end point. But the idea is, if your computer has a container interface, a storage interface, and all of that why not a standard “API” for AI agents on the desktop?

My initial ideas were combining linux-mcp-server and goose into a supplementary UX. We’re doing this now with troubleshooting and it works awesome and has caught (and helped fix!) bugs. (Alpha state)

I’m doing everyday things already (not on local yet but it’s only a matter of time) - you know how GNOME comes with those nerdy names for sound sources by default?

image
image791×286 17.4 KB

I did “Rename all my sound devices to make sense and delete spidf no one uses that.” Now I have “Headphones”, “Speakers”, and “Microphone” and it’s perfect.

All of that config is just flipping dconf keys, we just need to put the structures in place for them to be able to drive but done properly and safely.

My end goal is to speak into my mic: “I’m back from a long holiday, install all the cool apps from flathub from the past month” and then bazaar opens up with the right categories, etc. I don’t think you’re crazy at all of this is just basic Star Trek lol. “Bluefin, turn on developer mode” and then it’s like Iron Man turning on his hud, wallpaper changes, some kind of cool “snikt” sound comes out. All of my terminals, IDEs, agents all automatic. I’m all in man.

Other Update from midweek - lots of “it should have done this but it didn’t” in the factory, we’re looking for something that enforces contracts. jbeda’s working on a tool for this: