Justy: Cody, this Red Hat thing matters because agentic coding is leaving the toy demo phase. People are letting agents touch real repos now.
Cody: Yeah, and the messy part is the laptop. The agent is sitting next to secrets, local containers, test clusters, half-finished branches. That is where the nice slide deck gets awkward.
Justy: I'm grabbing the cold coffee from your counter, by the way. DC humidity is doing that thing where it feels personal. Anyway, this Red Hat release feels like it is aimed at the team lead who already has OpenShift, already has compliance reviews, and does not want a whole new AI island.
Cody: My kitchen has become a weather station with mugs. But yeah, the core move is: bring agentic development into the stack those teams already operate. Red Hat Desktop is generally available, and the Red Hat build of Podman Desktop now has commercial support.
Justy: That support angle is not glamorous, but it is the product. A developer can run containers on Linux, macOS, or Windows, build locally, then connect to local or remote OpenShift clusters for unit testing. For enterprise buyers, boring support is often the thing that lets a pilot become standard.
Cody: Right.
Justy: And the adoption barrier is not, do developers want help from agents? I think most do. It is, can we let this thing operate near production-shaped code without creating a governance headache?
Cody: The clever bit is the local sandboxing for AI agents. Red Hat is saying you can build and test agents on your own hardware, but isolate their actions so an agent does not accidentally mess with the host operating system.
Justy: Mm-hm.
Cody: Then they tie that to Red Hat Hardened Images and Red Hat Trusted Libraries. The images are stripped down and scanned. The Python packages come with software bills of materials and cryptographic signatures, built around OpenSSF practices. So the agent is not just pulling random stuff because it sounded useful.
Justy: The sandbox piece is the part I would demo to a skeptical manager. Like, look, the agent can be helpful, but it cannot crawl across your laptop like a raccoon in the pantry.
Cody: That is exactly the fear, though. Agents make mistakes at machine speed. If the environment is disposable, signed, and close to production shape, you get more signal from experiments without handing the agent the whole house.
Justy: Red Hat also expanded OpenShift Dev Spaces, which is their secure, zero-configuration cloud dev environment. It already works with tools like Claude CLI, Microsoft Copilot, Cline, Continue, and Roo, and now there is a technical preview integration with AWS Kiro.
Cody: Oh interesting.
Justy: That user story is pretty clear. A developer keeps their preferred coding assistant, uses a cloud IDE, and can point at frontier models or private models. Red Hat is trying to be the governed workbench around the tools, not the only tool.
Cody: That is the contrast with the hyperscaler agent platforms. Gemini Enterprise, Bedrock AgentCore, Copilot Studio, those are more managed-service shaped. Easier in some ways. But you trade away some architectural control, especially if your estate is hybrid and old in places.
Justy: Sure.
Cody: The part I find genuinely useful is the AI skills repository. These skill packs give agents step-by-step know-how for Red Hat products, like OpenShift, OpenShift Virtualization, and SRE-style security reliability work. And because they connect through MCP servers, the agent can reach outside systems without every team inventing a custom integration.
Justy: My only hesitation is maintenance. Skills sound great until nobody owns the skill file six months later. If they are versioned, reviewed, and evaluated like code, great. If they become a folder of magic prompts, people will quietly stop trusting them.
Cody: Right, right.
Cody: That is why the trusted software factory preview matters. It is a CI/CD setup based on CNCF practices and Red Hat's own build processes, something teams can use as-is or copy and modify. They are basically trying to make agent behavior auditable, not mystical.
Justy: Then there is Fedora Hummingbird Linux, which is the more experimental lane. Free rolling release, fast updates from upstream communities, anonymous agent-driven pulls, SBOMs, and packages that are supposed to be clear of known CVEs.
Cody: I would treat Hummingbird as the stage-two proof-of-concept box. Not where your stable production estate lives. More like, give agents a fast-moving OS that matches open source velocity, while RHEL stays the slow foundation.
Justy: That split makes sense for the market. Regulated teams, private cloud teams, companies with long-lived infrastructure. They want AI productivity, but the user story is, do not make me abandon the architecture I am paid to keep under control.
Cody: For a weekend test, I would keep it small. Install Podman Desktop, run `podman machine init` and `podman machine start`, then create a tiny MCP server that exposes read-only access to sample logs. Put the agent in a container, give it a Red Hat or Kubernetes troubleshooting skill, and see if it can summarize failures without touching the host.
Justy: Solo builder version: make a fake support queue, a local container app, and a little skills repo in Git. The win is not a flashy agent. It is proving you can inspect what it knew, what it did, and what changed.
Cody: Which is less cinematic than an autonomous coding army, but a lot closer to something a company would actually approve.
Justy: Exactly. Alright, Cody, finish your coffee. Episode 401 somehow became the sensible agent one.