As AI Work Cards Travel Between Tools, Review and Approval Become the Service

When a way of working became a file

Anthropic announced on October 22, 2026, that it would donate Agent Skills to the Agentic AI Foundation, a Linux Foundation project focused on agentic AI. Agent Skills is a format that packages work steps, cautions, reference documents, and required executable files for an AI assistant.

At its core is a text file called SKILL.md. Write the card’s name, purpose, and required procedure in this file, then place templates and reference materials in the same folder. An AI assistant that can carry out work steps reads them when needed.

The assistant does not read every card in full from the start. It first looks only at each card’s name and short description, then loads the detailed procedure when it matches the assigned task. This is meant to let it select only what it needs even when there are hundreds of work cards.

Anthropic said that more than 70 tools use the format, with 1 million public cards and roughly 280 million installations. These are the company’s own figures, and it has not disclosed how duplicates were removed. Still, they signal a supply large enough that this is no longer easy to view as a feature of one company’s product.

Making cards public does not make them safe. Security company Snyk said it inspected 3,984 cards across two public marketplaces and found serious security issues in 534 of them. It said 76 were malicious cards intended to steal passwords, leak information, or do similar harm. Anthropic also advises people not to install cards from untrusted sources, and to inspect the instructions, executable files, and external connections inside a card.

Before more cards, people need a reason to trust them

Consider the owner of a 12-person window-installation business. The owner copies addresses and window dimensions received through phone calls and KakaoTalk into estimate files. Site photos go into a group chat, visit dates go on a calendar, and pre-installation checks are carried again as paper printouts.

Copying the same customer information across three or four places can lead to a wrong unit number or measurement. An experienced employee knows what else to inspect on an old window frame. A new employee can miss it even after looking at the photos, so the owner checks again before sending an estimate.

A work card for this business could contain steps such as: “separate the inquiry into address, dimensions, and preferred date”; “ask for items to verify in photos”; “do not create an estimate if information is missing”; and “get the owner’s approval for the final amount.” Frequently used estimate language and a site-check list can sit in the same card.

With the card applied, an employee can enter a customer conversation once and receive missing questions, a site-check list, and an estimate draft on one screen. After a visit, they can add photos and have the card flag missed checks, then turn only approved estimates into text to send to the customer.

The advantage of an open format is that it does not lock this procedure into one AI service. If the price of today’s tool rises or its company changes policy, there is a possibility of moving the card and its reference materials to another tool that reads the same format.

Physical measurement, judging window-frame condition, safety decisions, discount approval, and promises to customers still remain with people. A work card is not a tool for removing an experienced worker’s judgment. It is closer to a device for preventing repeated questions and records around that judgment from being missed.

If this effort fails, it will likely be because someone made an attractive card without including real exceptions, an outdated card kept following a changed procedure, or a risky card was given access to customer files. That is why the first services to sell may be less about marketplaces with many cards and more about showing whether a card worked in a company’s real cases, who approved it, and when it was changed.

Outside the card-writing task, the focus is testing and control

Glean bundled account planning, customer-problem classification, and content-writing procedures aligned with brand standards into work cards. In a Glean case published by OpenAI, evaluation accuracy for Salesforce-related work rose from 73% to 85%, while time to first response fell by 18.1%.

These figures come from supplier-side material and need independent verification. But they show what to measure. The reported performance indicators were accuracy and time to first response, not the number of cards created.

Google Cloud operates a marketplace where companies can find and deploy AI assistants for work. Suppliers can choose among free, monthly subscription, usage-based, and combined base-fee-plus-usage pricing. Company administrators can expose only approved assistants in the internal catalog.

Google said the marketplace has thousands of pre-validated assistants, but it did not disclose payment conversion rates or revenue for a particular agent marketplace. What matters is the structure: rather than letting staff install directly from a public marketplace, it lets a company administrator review and allow what employees can see.

Four things to build from here

1. A studio that turns field interviews into work cards

This is a service that interviews an owner or experienced worker about their usual sequence and exceptions, then creates both a card that an AI assistant can read and a checklist for employees. It is for window installation, signage, and equipment-installation businesses where one site manager’s memory holds the methods for estimates, ordering, and inspection.

Now that a shared format creates a way to test a card across several tools, this can start now. The first screen should have a button labeled “Load one real job handled recently,” plus four fields: questions, judgments, approvals, and exceptions.

2. A test lab that repeatedly runs cards against real cases

This is a service that reruns past cases before and after a work card changes, finding omissions, wrong answers, and prohibited actions. It is for internal operators at labor consultant and tax accountant offices, where consultation standards or filing deadlines change often and a person must check an answer before it is sent.

Once cards can move between tools, teams need to check whether they produce the same result in each tool. The first screen should prominently show a list of anonymized past cases and counts for “Pass,” “Human review needed,” and “Fail.”

3. A control room for card approvals and change history

This is a service that records who created a draft, what a field manager changed, and which version employees are using. It is for operations teams at food-service brands where headquarters sets cooking, refund, and hygiene procedures and multiple franchise locations follow them.

As work cards become easy to copy as files, the problem of old cards remaining at individual locations grows. The first screen should show the cards currently in use, cards awaiting approval, locations that have not switched, and the name of the last approver.

4. An installation gate that filters risky external cards first

This is a service that explains in plain language whether a card reads customer lists or computer files, sends information to an external address, or contains hidden executable files. It is for managers at small clinics, private academies, and real-estate brokerage offices that handle customer personal information but have no dedicated IT staff.

As public cards increase, it becomes harder for people to distinguish useful cards from malicious ones by sight alone. The first screen should group requested permissions into “Read documents,” “Send externally,” and “Execute files.” If a risky item is present, it should request approval from a responsible person instead of proceeding with installation.

Why this matters where you are

You can check whether teams in your market repeat the same multi-step work, make avoidable omissions, and need a person to approve high-stakes decisions. The tools, marketplaces, and rules may differ, but portable work procedures create the same need to test what happens when a procedure runs in another tool. Start by measuring accuracy, response time, missed steps, and approval history in one real workflow.

What to check today

Call three businesses in the same industry with between five and 20 employees for 10 minutes each. Ask about one task they handled at least three times in the past month using the same sequence. If two of the three describe a similar process with five or more steps and can recall at least one point they had to recheck because something was missed, that work is worth testing with inspection and approval features before card production itself.