Cloudflare Agent Skills become files, dropping the build step
Agents SDK skills now load straight from a Workspace, R2 bucket or code, so an agent can pick up a new SKILL.md with no build or redeploy. The old compiled API is gone.
Cloudflare changed how Agent Skills work in its Agents SDK on 29 October. A skill is now just a folder with a SKILL.md inside, dropped into an agent’s Computer Workspace under /workspace/.agents/skills/. The agent picks it up on its next turn, with no build step and no redeploy, and can write its own skills the same way.
Skills can load from three places:
getSkills() {
return [
skills.workspace(this.workspace), // /workspace/.agents/skills
skills.r2(this.env.SKILLS_BUCKET, { prefix: "skills/" }),
];
}
Each SKILL.md is checked against the Agent Skills specification; one that fails validation still loads, with a warning logged instead of a hard failure. This needs Agents SDK 0.32.0 and Think 0.23.0 or later, and it is a breaking change: SkillRegistry, the agents/skills/compile step, the agents:skills Vite import, and Think’s skillWorkspace are all gone, and runner() has moved to agents/skills/runner.
Why it matters: if you’re building agents on Cloudflare’s stack, skills stop being something you compile and ship and become something you can edit, or let the agent edit, while it’s running. That’s useful for iterating on agent behaviour without a deploy cycle, but anything importing the old compiled API will break on upgrade.
The caveat: this is the second breaking change to how Think and the Agents SDK store agent state in two weeks, after the @cloudflare/computer workspace migration on 20 October. If you’re on that stack, budget time to check both.