The short version

  • Package a defined method with its supporting material.
  • Load the context the task needs, with clear source and version information.
  • Enforce permissions and approval at the tool boundary.
  • Review the complete package and retest it when anything changes.

Package the method

On 22 October 2026, the Agentic AI Foundation announced Agent Skills, Open Responses, OpenEnv and OpenSandbox as new projects under its governance.1 The announcement brings related work on capabilities, interfaces, evaluation environments and execution together. It does not mean agent skills were invented that week, or that every implementation behaves identically.

The Agent Skills format combines a SKILL.md file with optional scripts, references and assets.2 A useful package might describe how to check a document, include a schema for the result and supply a script for a repeatable calculation. Its description should say when the method applies and what a completed task looks like.

Load the right context

Skills use progressive loading: an agent can discover a short description, read the instructions when relevant and open supporting resources as needed.3 This helps keep unrelated material out of the working context. It also makes the description consequential: an unclear description can cause the agent to miss the right method or select the wrong one.

Keep the procedure concise and the references identifiable. State which inputs are required, which source governs a decision and what to do when evidence is missing. Record the skill version used for each task. Saved business records and approval decisions need their own durable storage, with appropriate access and retention controls.

Keep authority explicit

Consider an illustrative skill for preparing a maintenance request from an approved procedure. It can explain how to find the asset reference, identify required fields and present a draft. This is a design example, not a reported client deployment.

The connected tools must still check the caller’s access, validate the asset and destination, and require the appropriate approval before creating a request. A sentence inside the skill cannot grant a broader permission. Neither can an instruction found inside the procedure being read.

Separate the responsibilities: the skill describes the method; the model proposes the next step; the application enforces the permitted operation and records its result. If the procedure changes, the draft may need another review. If a reply is lost, the application must reconcile the operation before creating a replacement.

Control the execution

A package containing scripts brings executable behaviour into the design. OpenAI’s current skills guidance calls for reviewing both instructions and supporting files, and distinguishes local execution from hosted environments.4 Read the actual package and its dependencies before making it available to an agent.

Define the files, credentials, destinations and network routes the execution environment can reach. Set limits for time, output and resource use. An isolated environment reduces exposure only within the boundaries it enforces; it does not prove that its contents are trustworthy or its outputs are correct.

In the maintenance example, a formatting script may need a temporary draft file. It does not automatically need the service account that creates work requests. Keep those capabilities separate, and decide how temporary files and generated records are retained or removed.

Release a known version

A skill update can change behaviour as materially as a model update. Review the instructions, scripts, dependencies and destination references together. Keep the package’s origin, maintainer and approved version visible. A familiar package name should not conceal a changed executable or new network dependency.

  1. Compare the change. Record what changed and which tasks or permissions it affects.
  2. Exercise the boundaries. Include missing evidence, denied access, altered instructions and interrupted actions.
  3. Check completed work. Inspect outputs and resulting system state, alongside response time and cost.
  4. Control the rollout. Retain the prior approved package where supported, and define when to stop or recover.

Reuse can reduce repeated implementation work, but a shared format does not make platforms interchangeable. Tool availability, instruction handling, model behaviour and execution controls can differ. Test the skill in the runtime where it will actually operate, including after a platform or model change.

Leave an inspectable handover

The useful deliverable is a defined method, its reviewed package, scoped connections and evidence of how it behaves. Give the owner a clear way to update it, investigate a failed task and stop further action.

Sources and further reading

  1. Agentic AI Foundation announces four new projectsThe Linux Foundation · 22 October 2026↩
  2. Agent Skills specificationAgent Skills↩
  3. Adding skills support to an agentAgent Skills↩
  4. Skills: versioning, execution and safetyOpenAI documentation↩