Ubuntu 26.04 is the OS of the AI agent era, and Canonical’s Mark Shuttleworth explains why:

In his speech at the Ubuntu Summit 26.04, Shuttleworth presented open source as the “raw material” for the next wave of technological disruption. Specifically, he said the pace of AI-driven software innovation has outpaced traditional packaging and release processes. For AI, Linux users need to move beyond Advanced Packaging Tool (APT) and Red Hat Package Manager (RPM) to signed, auto-updated, rules-based snaps. Of course, snapshots have long been Canonical's answer to upstream software delivery, but now AI requires updates at internet speed without sacrificing auditability or control, and that means snapshots.

Shuttleworth cited new telemetry from developer Alan Pope's Snap Store dashboard that shows dozens of snap updates arriving in a single morning, on architectures ranging from x86 and Arm to RISC-V and Power, all coming from the same bits tested. He positioned snaps, with containment, phased deployments, channels and enterprise management, as "the best and safest way to deliver bits to any Linux distribution on the planet."

While Shuttleworth defended snaps in general, Ubuntu VP of Engineering Jon Seager explored a new user-facing behavior: finer-grained permission prompts for snapped apps, similar to those in Android and iOS. For example, when a confined application first attempts to access the camera, the desktop can now display a prompt asking the user to grant or deny access, thanks to new plumbing from the kernel and AppArmor via Snapd and GNOME's display manager.

Sandbox everything: from snaps to LXD, Multipass and microVM

The other reason Ubuntu is the operating system you want for AI, according to Shuttleworth, is security. With this version of Ubuntu, everything can work in a layered toolbox. All? All. These are not just apps, but also AI agents and third-party software development kits (SDK). On Ubuntu today, this covers instant containment, Docker/OCI containers, LXD system containers, traditional virtual machines (VMs) through Multipass, and a new generation of microVMs that blur the line between containers and virtualization.

This combination is essential, Shuttleworth says, for "agentic engineering," where organizations may want to run thousands of agents, each thinking it has a full Linux system while actually being tightly constrained in terms of density and security. LXD-based system containers give the illusion of complete machines for agents, while microVMs, delivered via an "Open Shell" snap-in that runs agent-hardened environments for tools like Claude or Copilot, add hardware-enforced isolation when a kernel boundary is not sufficient.

Workshop: A new way to integrate developers and agents

Also, a new concrete piece is Workshop. This is a Canonical tool built on top of LXD for creating “agent workspaces”. It aims to solve a long-standing problem: combining sensitive developer credentials with untrusted or semi-trusted code.

Developers or teams can commit a workshop definition to a repository. Thus, the integration of a new human or a new agent becomes “a git clone, a workshop launch”. With these, the company claims, you can launch sandboxed development environments and composable, repeatable agent workflows with a single command, while keeping your host system isolated.

Workshop works by starting a system container and then selectively binding high-value secrets and resources, such as SSH keys for signed commits, access to specific datasets, and routes to remote Git servers, without dumping a developer's entire portable environment into the sandbox. Canonical already works with ISVs to ship signed SDKs to a dedicated Workshop store so that SDKs and closed agents can work with Ubuntu and Debian packages in a controlled environment.

AI, Ubuntu and the challenge of “implicit features”

Seager picked up where Shuttleworth left off, arguing that Canonical had no choice but to be "at the heart" of AI and agents if open source was to have any say in the evolution of these systems. Rather than rushing to integrate an LLM gadget into the shell, Seager presented a two-track strategy: implicit AI features that quietly enhance existing capabilities, and explicit AI features that Canonical will deploy more cautiously.

On the implicit side, he highlighted accessibility and media as near-term opportunities: local text-to-speech, better camera autofocus, and microphone improvement powered by small on-device models that can run even on laptops with only a CPU. On the explicit side, he outlined a goal for Ubuntu 26.10: a desktop where "you can press a button and speak in any field you could type before," supported by templates like Whisper and built into every text entry surface in the system.

Seager made it clear that AI-driven accessibility is a primary design goal, not an afterthought. He called today's Linux screen readers "frankly rubbish" and argued that introducing a framebuffer or camera capture to an LLM could radically improve both the description of on-screen content and the presentation of possible actions to visually impaired users.

Beyond accessibility, Seager presented "new ways to interact with your machine" that build on Ubuntu's existing history of containment: in an agentic desktop, each tool an agent can invoke would be packaged as its own contained snap, giving fine-grained control over what the agent can do on the user's behalf. He promised something concrete to "play with in the next six months", describing it as a way for non-experts to get "20 years of Linux desktop hacking capability" through agents, without needing hacking experience themselves.

Additionally, on the AI ​​and HPC front, Seager highlighted Canonical's work with NVIDIA and AMD to make enabling GPU boring... in a good way. Ubuntu users can now "simply install CUDA and apt install ROCm", with Canonical and vendors working together to ensure drivers and stacks are properly integrated and tested on 04.26.

Seager added that his own AMD GPU "never sang as well as 26.04" and that, for the first time, he "didn't have to endure any pain" to get there. In combination with Ubuntu's work on architecture variants, providing entire archives compiled for specific instruction set levels such as amd64v3, Canonical wants to ensure that the expensive acceleration hardware that companies purchase is fully supported by Ubuntu Linux and its bundled tools.

Keeping Ubuntu accessible in a token-limited world

Both Shuttleworth and Seager vowed to keep the historic promise of Ubuntu alive, shipping "precisely the same bits" to hedge funds and suburban Calcutta kids, in a world where AI use is measured in expensive tokens. Shuttleworth warned that tying productivity and even basic understanding of code to proprietary models hosted in the cloud risks excluding the "poorest members of our digital society" unless open models and open tools remain a priority.

Seager, for his part, rejected both “moral” disengagement from AI and vanity measures such as “who can spend the most tokens.” He argued that open source savvy players like Canonical need to stay engaged, help the community navigate a complicated period of AI-generated "ineffective contributions" and guide the eventual convergence to a new generation of high-quality open source components, now with agents and some AI in the toolbox.

Beyond AI: Rust, Security and Cryptography

Seager also highlighted how Ubuntu 26.04 integrates memory safety into the base system. He highlighted three pillars: critical utility rewrites based on Rust, a new cryptographic foundation based on Rust called Universal Public Key Infrastructure (UPKI), and a unified time synchronization stack based on Rust.

As of version 26.04 Long Term Support (LTS), coreutils such as mv, cp, rm and ls are now supported by the Rust-based uutils project, following two Canonical-funded security audits. Sudo has been replaced by sudo-rs, a Rust implementation that removes long-accumulated "ill-informed" features and enforces memory security at the edge of privilege on every Ubuntu machine. Next, Canonical plans to swap bzip2 for a Rust implementation that Seager says is "up to 50% more efficient," with Zlib and Zstandard targeted by 04.28 - changes that he says could result in significant global energy savings given the scale of use of those codecs.

The goal of UPKI is to bring browser-level PKI hygiene to the Linux command line. Today, Seager noted, curl happily ignores certificate revocation lists, and command-line TLS often breaks over misconfigured certificate chains that browsers silently tolerate. UPKI will centralize revocation, intermediate prefetching, and eventually post-quantum algorithms such as Merkle tree-based schemes, with glue code written for OpenSSL, GnuTLS, curl, and others so that tools in the stack can consume the same modern PKI data.

On the other hand, a new NTP-rs utility will provide NTP, NTS and PTP "in a single binary configuration", aiming to radically simplify precise timing configuration in Linux.

Accelerating Ubuntu releases in the AI era

Seager described the 26.04 release as the first LTS delivered under a new engineering "manifesto" that included monthly shipping discipline enforced by an entirely new release pipeline built with Go and Temporal. The team, he said, hit every monthly goal, which made the LTS release smoother.

He added that Canonical has also been quietly rebuilding its community and communications capabilities. Seager claimed that Canonical has added more core developers in the last six months than in the previous three years and has deliberately increased its blogs, Mastodon posts, podcasts and community appearances. The result, Seager joked, is that "for anyone who doesn't like Ubuntu, it's a bit of a tough time... you literally can't escape us on the Internet."

This also means that Ubuntu can keep up with the incredible pace of AI.