Agent Skills (Preview)

Messages and traces describe what happened in particular interactions. Reusing that material in another task requires selecting relevant evidence and distinguishing a repeatable procedure from incidental details.

Distillation derives a portable SKILL.md package from a scoped set of memory. This is different from copying an entire workspace into a prompt or treating every successful trace as a general procedure.

Creation selects and snapshots a scope, synthesizes a procedure, checks its grounding, and packages the result. Review decides whether that result should be published. Use means loading the resulting package into an agent capable of interpreting it.

These are separate responsibilities. Publishing a package is not the same as executing its steps, and downloading a ZIP is not evidence that an agent has loaded it. The lifecycle and concrete operations belong in Skills API.

Provenance links connect generated claims and steps to their source records, including :GROUNDED_IN relationships in the documented service model. A reviewer can inspect what evidence supports a procedure and whether the interpretation matches that evidence.

Grounding is not a guarantee of truth. A source may be wrong, a record may omit an important condition, or the synthesized procedure may overgeneralize one execution. Quality gates and human review address these concerns, and a low-coverage distillation can be withheld rather than published.

A skill is derived from memory rather than a fourth memory layer. Messages supply context, entities supply domain information, and recorded traces supply actions and outcomes. An ontology class can help define a narrower scope.

A recorded tool call describes one action in one environment. Turning it into a reusable step requires preserving its prerequisites and checking whether its assumptions still apply. The source records remain relevant even after a package is published.

The documented bundle contains a SKILL.md, procedure material, provenance, and supporting references. A published package’s attestation can help a consumer verify its integrity and signer.

Integrity and provenance answer different questions from factual accuracy or suitability for a new environment. The consumer still needs to review the procedure and provide any required tools or permissions. See Package and attestation details.

The service’s skill feature is in preview. Distillation, composition, and execution have distinct enablement and limits; a capability described by the service is not necessarily enabled for every deployment.

The documented execution path is gated dry-run planning, not autonomous tool execution. Running a downloaded procedure is the receiving agent application’s responsibility. Scope a distillation around one coherent procedure rather than assuming that a workspace containing unrelated tasks yields one useful skill.

Follow the skills walkthrough for its demonstrated scope, real identifiers, and verification outcome.