An AI assistant can produce a polished report and still miss what your team actually needs. It might use the wrong reporting period, overlook an internal definition or present an estimate as a confirmed result.
Correcting those mistakes often means repeating the same instructions every week.
Agent Skills offer a way to make those instructions reusable. They package a task’s methods and supporting resources so a compatible AI agent can consult them when needed.
On October 22, 2026, the Agentic AI Foundation announced that Agent Skills, contributed by Anthropic, was joining its portfolio of projects. The announcement brings renewed attention to a practical question: how can organizations give AI access to their working methods without rebuilding those methods for every tool?
What are Agent Skills?
Agent Skills is an open format for organizing instructions and resources around a specific task.
At its core, a skill is a folder containing a file called SKILL.md. That file identifies the skill, explains when it should be used and provides the instructions an agent should follow. The folder can also include templates, reference documents and executable scripts.
For a business team, that could mean a reusable procedure for preparing a client update, checking a spreadsheet or applying editorial guidelines.
“Teaching” an agent in this context means supplying guidance it can read while working. Installing a skill does not, by itself, retrain the underlying model or permanently change its knowledge.
That distinction matters. The quality of the outcome still depends on the model, the information available and the tools it can use. Our analysis of Google’s Gemini agent for workplace tasks explores how those elements come together when an assistant is expected to complete work across applications.
What changed with the foundation announcement?
Agent Skills existed before the October announcement. Anthropic introduced the approach in October 2025 and subsequently published it as an open standard in December 2025.
The latest development concerns the standard’s governance. Agent Skills is now hosted by the Agentic AI Foundation, with its direction overseen through a committee involving maintainers from multiple organizations.
For organizations considering skills, this creates a reason to evaluate the format as shared infrastructure. A documented procedure may remain useful across compatible agents as tools change.
However, a shared format does not guarantee identical behavior. Moving instructions between products is only part of moving a working process.
How a skill works
The official documentation describes a process in which agents load information in stages.
First, the agent reads the names and descriptions of available skills. When a description matches the task, it can load the full instructions. It then consults supporting files or runs included scripts as needed.
This approach avoids loading every procedure and reference document into every conversation. It also illustrates why files available to an AI and information actually used in its response are different things.
The description therefore has a practical role: it helps the agent decide when the skill is relevant.
A vague description such as “business reporting” provides little guidance. A description explaining that the skill prepares weekly sales summaries from approved exports gives the agent a clearer trigger. The specification explicitly calls for descriptions covering both what a skill does and when to use it.
A practical example: the weekly sales report
Consider an illustrative sales team that produces a management update every Monday.
Its report needs to cover confirmed revenue, changes in the pipeline and opportunities requiring attention. The team also has specific rules: exclude cancelled orders, separate forecast revenue from closed sales and identify missing data.
A skill for that workflow could bring together:
| Component | Example content |
|---|---|
| Task instructions | Which period to report on and which records to include |
| Business definitions | What counts as confirmed revenue or an active opportunity |
| Report template | The sections and tables managers expect |
| Worked example | An approved report using fictional data |
| Verification steps | Check totals, flag missing fields and identify source files |
The user could then request the weekly update without restating the entire procedure.
The important test would be whether the agent applies those rules correctly when the data becomes awkward. What happens when a cancellation appears after the reporting period? What if a source file is missing? Does the agent ask for clarification or silently invent an answer?
Those cases reveal whether the instructions are useful beyond a clean demonstration.
What can move between AI tools?
The standard gives teams a common way to package guidance. Its specification also allows authors to record compatibility requirements, such as required software or network access. Supported scripting languages depend on the agent implementation.
That leaves several questions to check before moving a skill:
- Can the destination agent read the included files?
- Can it run any required scripts?
- Does it have access to the relevant business systems?
- Does it interpret the instructions well enough to meet the team’s quality requirements?
The AAIF itself notes that agents still need the appropriate capabilities and access to perform the task.
For a team, the practical implication is to test the workflow again whenever the agent, its tools or its permissions change. Reusing the folder can reduce setup work; it does not remove the need to validate the result.
Skills need maintenance and review
A reusable procedure can spread a good practice across a team. It can also spread an outdated assumption.
A reporting skill that contains last year’s revenue definition may produce consistent but misleading results. A writing skill might continue applying a brand policy that has already changed.
A sensible operating approach is to assign an owner to each important skill, record revisions and review it when the underlying process changes.
Security also matters because skills can contain executable code and instructions that influence an agent’s actions. Anthropic warns that malicious skills may cause unintended behavior or data exposure, and recommends checking their contents, dependencies and external connections before use.
For a business workflow, review should therefore cover both the procedure and what it enables the agent to do. Demystia’s guide to permissions for AI agents handling emails, files and purchases provides practical examples of how to define those boundaries.
How to judge whether a skill helps
Demystia’s suggested starting point is one recurring task with a result that a person can assess.
Run a small comparison using the same inputs, model and available tools, with and without the skill. Include straightforward cases and difficult ones. Repeat the exercise enough to see whether improvements are consistent.
Useful measures include:
- How many factual or calculation errors remain?
- How often are required sections or checks missed?
- How much time does a person spend correcting the output?
- Does the agent identify missing information appropriately?
- Does the overall task take less time or cost less?
The most useful outcome is a measurable reduction in correction work while maintaining quality. A more polished first draft, on its own, is weak evidence.
That is why evaluation should include the people checking the results. Our analysis of AI agents and the human attention bottleneck examines how review effort can limit the value of faster automated production.
Making working methods reusable
Agent Skills make a familiar management task relevant to AI: documenting how good work gets done.
For Demystia readers, the opportunity is to capture a procedure that already works, give it clear examples and test whether an agent can follow it reliably.
The foundation announcement strengthens the case for a shared format. The value for an individual team will depend on the quality of its instructions, the suitability of its tools and its willingness to maintain both as the work changes.
Sources
- Linux Foundation — Agentic AI Foundation Announces Four New Projects, October 22, 2026.
- Agentic AI Foundation — Agent Skills: Sharing Team Expertise Across AI Agents, October 22, 2026.
- Agent Skills — Overview.
- Agent Skills — Format Specification.
- Anthropic — Equipping Agents for the Real World with Agent Skills, October 16, 2025, with a December 18, 2025 update.
Sources checked on October 25, 2026. The sales-report workflow is an illustrative example, not a reported deployment.