Agent Skills for DevOpsKubernetesCI/CDDockerPlatform EngIaCSRE
Build them. Eval them. Prove them.
A hands-on workshop for DevOps, platform, and software engineers. Learn to use, build, and manage agent skills for infrastructure work, then prove with evals in CI that they make your AI agents better, not worse. Tips for Claude Code and Codex, and it works with any agent that supports SKILL.md. About 2.5 hours, free thanks to Chainguard.
- Project
- Skills
- Rev
- 2026.12
- Drawn by
- B. Fisher
- Approved
- CI
Skills, Soup to Nuts
Everyone installs skills. Few measure them. We'll use them, build them, eval them, run those evals in CI, and manage them across a team.
Assembly sequence
Part 1 ยท Use and build
- 10m Skills vs. AGENTS.md vs. MCP vs. subagents vs. hooks vs. plugins
- 20m Using skills in daily ops work, with Claude Code and Codex tips
- 30m Building best-practice skills for infrastructure, live
- 20m Catch-up skills: teaching models what's new
Part 2 ยท Prove and scale
- 40m Evals: baselines, graders, fixtures, and variance
- 15m Running skill evals in CI without blowing the budget
- 15m Team skills: plugins, marketplaces, versioning, supply chain
- 10m Q&A and your take-home resources
Run the same tasks with and without a skill. Programmatic graders first, LLM judges only where you need them. Get a number, not a vibe.
My technique for teaching a model what shipped after its training cutoff, like the newest Kubernetes features, so it stops writing last year's YAML.
Gate skill changes on pull requests so a bad edit can't quietly make your agent worse, and keep the token bill sane.
Plugins vs. skills vs. MCP, marketplaces, versioning, and a skill supply chain as tight as your container supply chain.
Same Task. With and Without the Skill.
A skill changes how your agent behaves in your infrastructure. A good one makes it nail the Dockerfile it used to botch. A bad one quietly makes things worse. Here's what my own open source skills measured, and you'll learn to build the harness behind these numbers.
Score = share of eval assertions passed on the same tasks, same grader. Preliminary results; full tables and caveats in github.com/BretFisher/skills.
Where Do Your Skills Come From?
Public skill repos are untrusted code that runs with your agent's permissions. In the workshop I'll show what skill supply chain risk looks like, and how Chainguard Agent Skills fits into a team's setup.
Hardened, not just scanned
1,000+ popular community skills, reviewed and rewritten to remove risks like over-broad
permissions and credential harvesting. Each ships a HARDENING.md audit log.
Kept in sync with upstream
When the original skill changes, Chainguard re-hardens it automatically, so you get updates without giving up review.
Signed and pinned
Hardened skills are Sigstore-signed and verifiable, with version pinning and rollback, for every agent your team uses.
A registry for your own skills
Serve hardened community skills and your team's internal skills from one private registry, instead of copying files between repos.
Bret Fisher
For a decade, my courses, videos, workshops, and talks on containers and DevOps have reached more than a million people. These days I spend most of my time on one question: how do we make non-deterministic AI agents safe and reliable enough to automate infrastructure?
Skills turned out to be a big part of the answer. I build and maintain open source DevOps skills with their own eval harness, and I teach skills and evals in the Agentic DevOps Guild, where 50+ teams meet weekly to work through this stuff together.
Save Your Free Seat
Watch the Workshop Replay
Live online, Thursday, December 10, 2026, 12 PM ET. Can't make it? Sign up anyway and I'll email you the recording, the skills, and the eval tool after the live event.
Sign up and I'll email you the full workshop recording, the skills, and the eval tool.
Submitting this form will cause you to get emails from Bret Fisher and Chainguard. Unsubscribe anytime.