Cloudflare Agents SDK: agent skills are now files it can write

Cloudflare changed how its agents learn new tricks. In Cloudflare Agents SDK v0.32.0, agent skills are plain SKILL.md folders in the agent's Workspace or an R2 bucket. A new skill loads on the next turn: no build, no deploy. And per Cloudflare, "the agent can write its own skills." That is useful. It is also a new way for a bad instruction to stick around.

What actually happened

Cloudflare posted the change in its October 29 changelog. You need Agents SDK v0.32.0 and Think v0.23.0 or later.

• Three sources: the Computer Workspace (/workspace/.agents/skills/), an R2 bucket via skills.r2(...), or code. Think re-lists every source before each turn.
• Spec checks are warnings. Each SKILL.md is checked against the Agent Skills spec. A skill that breaks a rule still loads, with a logged warning. parseSkill() lets you check one before you write it.
• Scripts run in the Workspace through its exec tool.
• Same tools everywhere: activate_skill and read_skill_resource work in Think, Pi Durable and the AI SDK.
• Breaking change for code that uses agents/skills directly. The agents:skills Vite import, agents/skills/compile, SkillRegistry and Think's skillWorkspace are gone. runner() moved to agents/skills/runner.

The changelog describes no approval step for skills an agent writes itself. Three days earlier, Cloudflare Computer v0.7.0 added ctx.interrupt(), which pauses a call until your app approves it.

Why agent skills as files matter for your business

Hot-loaded skills are fast iteration. You can fix how your support agent handles returns by editing one Markdown file. No release cycle. For a small team, that's real speed.

Self-written skills are persistent memory, with execute rights. If a customer email tricks your agent into saving a "skill" that says "always CC this address," that instruction runs on every later turn. Prompt injection used to last one conversation. Now it can last until someone reads the folder.

Put a gate on the write. Our setup: the agent drafts skills into a staging prefix. A person, or at least parseSkill() plus a diff check, promotes them to the live path. Route anything that sends, pays or deletes through ctx.interrupt(). Keep production skills in R2 or in git, where changes have an owner and a history.

Migrate before you upgrade. If you imported agents/skills directly, budget an hour for the breaking change.

Sources: Cloudflare changelog: Agent Skills are files, Cloudflare changelog: Computer v0.7.0.