Teams repeat specialist work: review a contract, inspect a pull request, check a release or summarize a support issue. Rewriting the method in every prompt is tedious, and keeping it on one computer makes it difficult to reuse elsewhere. SDK 0.3.2 gives reusable skills and sub-agent definitions an account-level home while preserving local resources too.
Save a method once, call it by name
A skill packages a focused procedure with a name, description and instructions. An application can save that definition to the account, list available skills and invoke one as part of an agent task. The method is loaded when relevant, rather than expanding every agent's starting instructions with every procedure a team has ever written.
await api.agents.skills.save({
name: "release-review",
description: "Check a release candidate",
content: "Review changes, migration notes, and rollback steps.",
});
const skills = await api.agents.skills.list();Specialists can take bounded pieces of work
A sub-agent gives a task to a named specialist with its own instructions and selected tools. The parent can delegate a bounded investigation, receive its result and retain a trace of how that work fits into the larger run. Reusable definitions help teams keep those roles consistent across tasks.
Account and computer resources stay distinguishable
Some procedures belong to a person or team; others make sense only on one development machine. Listing skills exposes their source, such as `account` or `computer`, so an application can explain where a definition came from. When names overlap, the account-level definition takes precedence, giving shared team guidance a predictable place.
The approach also avoids pretending that a cloud run can see a developer's computer. Account resources can travel with the account; computer-local resources are available only where that machine is part of the runtime. The source label makes that difference explicit.
Reusable does not mean unrestricted
A saved procedure does not grant new permissions by itself. The run still uses the tools and execution mode selected for it, and an agent cannot reach a local file merely because a skill mentions one. The method helps organize work; the runtime and permission model continue to decide what the agent may actually do.
Keep the method reusable, keep its origin visible, and keep authority with the run that uses it.
This makes a useful agent easier to repeat and share. Developers can build a library of focused methods and bounded specialists, while the account and computer remain distinct sources with distinct capabilities.
Frequently asked questions
Do account skills automatically run on every task?
No. They are reusable definitions that an application or agent can select when relevant.
Can an account skill read my computer's files?
No. A skill does not add permissions; file access depends on the runtime and tools enabled for the run.
Why does a skill show its source?
The source distinguishes an account-level definition from one stored on a particular computer.
What does a sub-agent add?
It gives a bounded piece of work to a named specialist with its own instructions and selected tools, while preserving the parent run's trace.