Anthropic's Agent Skills: the new open standard for AI agents in the enterprise

The enterprise artificial intelligence industry is experiencing a minor earthquake with the movement of Anthropic and its Agent Skills proposalFar from releasing another closed feature, the company has opted to publish an open specification that It allows any organization to define, share, and govern AI capabilities in a standardized way.This is especially relevant for European companies operating in regulated environments.

In practice, this means that AI assistants stop relying on improvised prompts and start working with structured, versionable, and auditable skill librarieswhich can be reused across multiple teams, applications, and vendors. For companies in Spain and the rest of Europe already testing AI agents in legal, finance, or customer service, this approach It promises more control, less "black magic," and a more orderly integration with its internal systems..

What is Agent Skills and why does it mark a turning point in enterprise AI?

Agent Skills is, in essence, a common framework for teaching AI agents very specific work tasksThe knowledge is packaged into independent modules. Each skill is a folder or package with step-by-step instructions, scripts, usage examples, and specific resources that tell models like Claude how to act in a given professional context: generating a financial report in accordance with regulations, preparing a presentation with the brand guidelines, or processing a reimbursement according to a company's policy.

Instead of the classic approach of "asking things" from the model with long prompts, organizations can create internal collections of skills that reflect their real processesThese libraries are shared across teams, reviewed as if they were code, and integrated into the tools already used daily. For many European companies, this approach better aligns with their needs for regulatory compliance, data governance, and traceability.

One important change is that Anthropic is not limited to using Agent Skills within its own ecosystem: The specification is published as an open standard.This is similar to what the company did with its Model Context Protocol (MCP), now widely adopted for connecting agents with external services. Any provider, whether a cloud giant or an industry-specific software company in the EU, can implement and extend the standard without being tied to a single vendor.

In a market where models from OpenAI, Google, Anthropic, and other players coexist, having a common language to describe agents' abilities It aims to reduce dependence on proprietary platforms and facilitate migrations or hybrid deployments, something increasingly valued by European banks, insurers or public administrations.

How Agent Skills work and what problem they solve

Agent Skills are presented as encapsulated modules that live between the language model and the internal systemsThe model is still the one that understands, reasons and converses, but when it has to "do" concrete things —check a balance, open a ticket in Jira, generate a regulatory report— it resorts to the appropriate skill, which precisely defines how to proceed.

Each skill usually includes a definition file (like the well-known SKILL.mdThis section describes, in a mixed format of YAML and structured text, the skill name, the steps to follow, the allowed parameters, usage examples, and the tools or APIs that can be invoked. No sensible steps are left to chance: They are implemented as deterministic code that calls business serviceswhile the model focuses on the conversational and decision-making aspects.

To improve efficiency, Anthropic has incorporated a design of “progressive disclosure”The assistant doesn't load all the details of every skill in context; it only accesses the complete information when it's actually needed. This way, an organization can maintain a very large library without overloading the model's memory, which is especially useful in complex environments like banks, telecoms, or large European retailers.

Another common component is the so-called orchestrating agent, that acts as a supervisor: receives the user's request, detects the intent, decides what combination of skills and tools is necessary and sequences themA simple billing query can trigger an intent clarification skill, an "explain my invoice" skill, and, underneath, a tool that queries billing systems without the user having to understand that complexity.

In this approach, skills become the fabric of execution of the agentsThe conversational level remains flexible, while procedures are defined, reusable, and subject to quality control. This It corrects one of the major shortcomings of the first AI-based bots and assistants, whose behavior was difficult to audit. and it changed unpredictably when the prompts were modified.

Openness, standard, and early adoption of the ecosystem

Anthropic's most striking move has been to publish the Agent Skills technical specification and its SDK as an open standard through agentskills.io, inviting the community and other providers to adopt and evolve it. This move follows the MCP, which has recently come under the management of the Linux Foundation within the Agentic AI Foundation, in which actors such as AWS, Google, Microsoft or Block participate.

Around Agent Skills, a early adoption by large technology companiesTools like Microsoft VS Code, GitHub, and coding agents such as Cursor and OpenCode have incorporated skills architecture to define development workflows. OpenAI itself has introduced very similar structures in ChatGPT and its developer CLI, with skill directories reminiscent of Anthropic's approach, suggesting a certain convergence within the industry toward this type of modularity.

Meanwhile, leading enterprise software companies —Atlassian, Figma, Stripe, Canva, Notion, Cloudflare, Zapier or RampCompanies like [company name] are publishing their own skills to connect their products with AI agents. These skills allow users to, for example, create tasks in Jira or Trello following internal conventions, apply brand styles to Figma designs, or automate marketing workflows without needing ad hoc integrations for each client.

The developer community is also getting involved: Anthropic's skills repository has accumulated tens of thousands of stars on GitHub and There are already thousands of publicly shared skills, ranging from utilities for manipulating PDFs to specific automations for engineering or financial teams.

This ecosystem is especially interesting for European companies that make intensive use of tools such as Atlassian, Microsoft 365 or Figma and want their AI agents to work with them while respecting internal policies, sector regulations and privacy requirements such as the GDPR. without relying on opaque extensions from a single provider.

From developer tool to enterprise infrastructure

When Anthropic introduced these capabilities in October, the skills were perceived mostly as A utility for developers and code enthusiastsThrough an interactive “skill-creator” in Claude, users themselves could generate the folder structure and the SKILL.md necessary to automate specific workflows, without major engineering deployments.

With the recent update, the company has shifted its focus to the enterprise: Agent Skills now integrates with organizational management toolsA central directory of skills and management functions designed for IT managers and security teams. The idea is for skills to move beyond scattered experiments and become stable, documented, and governed assets as part of the enterprise AI infrastructure.

In organizations subscribed to Claude's Team and Enterprise plans, skills can be managed from a central panelThis is where administrators decide which skills are provisioned to each user group, which are enabled by default, and which require opt-in. This layer of control allows for aligning agent usage with internal policies, which is crucial for highly regulated sectors in Europe, such as healthcare, insurance, and banking.

In addition, Anthropic has opened a Skill Directory of business partners It functions as a catalog of ready-to-use skills, with contributions from companies like Atlassian, Canva, Figma, Notion, Cloudflare, Stripe, Zapier, and Sentry. For many European SMEs and large companies, this type of repository streamlines pilot projects: instead of building everything from scratch, they can start with pre-tested skills and adapt them to their processes.

All of this suggests that, more than just a product feature, Agent Skills is evolving into a infrastructure layer on which to build AI agents and applications, in line with what the standardization of APIs meant at the time: a common language on which different tools can cooperate.

Productivity, use cases and benefits for European companies

The first real-world deployments show that the adoption of Agent Skills is not just theoretical. Engineering teams have reported productivity increases of up to 50%. thanks to the automation of repetitive tasks and the standardization of workflows such as code review, technical documentation, or test generation.

In the financial and accounting field, skills allow codify regulated proceduresFrom checks prior to issuing a report, to compliance controls that run automatically before approving certain transactions. For Spanish companies subject to European regulations—such as MiFID II for investment services or Solvency II for insurance—being able to translate these rules into auditable skills is an advantage over unstructured prompts.

In operations and back office, organizations are using skill libraries to sharing institutional knowledgeWhat was previously known only to a few veteran employees is now embodied in modules that an agent or a new worker can follow step by step, reducing dependence on specific people and accelerating internal training.

Even more ambitious experiments have been tested, such as Anthropic's internal project to manage a small merchandising store with agents equipped with skills in inventory, sales, and customer service. Although human supervision remained in some extreme cases, the tests suggest that Agents equipped with well-designed skills can execute end-to-end tasks in controlled environments.

In the European context, where the Commission and national regulators are beginning to demand greater transparency and control over AI systemsThis modular approach facilitates risk assessment: each skill can be documented, tested, and certified independently, while the overall model is used as a reasoning and natural language layer.

Risks, governance and skepticism surrounding the standard

Opening up Agent Skills is not without risks. By allowing anyone to post and share skills, There is a possibility that malicious or low-quality skills may emergewith instructions that could lead to errors, regulatory non-compliance, or even information leaks if connected to sensitive systems.

Anthropic advises companies that Limit the adoption of skills to audited sources and verified developersand that they integrate the review of these capabilities into their regular security and compliance processes. The company also participates in discussions with the community about who should manage the long-term evolution of the open protocol and how, an important issue if the standard is to be prevented from being captured by a single actor.

Another ongoing debate is the impact on the human skills within organizationsAs agents automate entire procedures, some experts warn of the risk of skills "atrophy": if a team gets used to AI always preparing reports, filing claims, or managing customer service processes, it may lose the dexterity to do it manually when something goes wrong.

Industry analysts also point out that, although MCP has become a de facto standard, It is not guaranteed that Agent Skills will repeat the same successOrganizations are already accustomed to working with standardized APIs and communication signatures, and there are multiple ways to teach capabilities to agents. In other words, the technical advantages of Agent Skills alone are not enough to ensure widespread adoption.

For European companies, accustomed to operating in multi-vendor ecosystems, this skepticism translates into caution: many are experimenting with Agent Skills in pilot projects, but maintaining in parallel strategies specific to orchestration and governance of agents, with layers of control that are above any specific standard.

Strategic advantages for founders and CTOs of startups in Spain and Europe

Beyond large corporations, Agent Skills opens an interesting window for European technology startups and scaleupsFor many founding teams, the real differentiator is no longer simply using the "best model" on the market, but codifying their own know-how in the form of proprietary skills that capture their processes, their way of working, and their understanding of the customer.

In this sense, investing resources in building libraries of skills that represent organizational intelligence This can become a long-term asset, comparable to owning a well-designed API or a robust data infrastructure. These skills can be deployed across different models and platforms, reducing dependence on a specific vendor and facilitating compliance with European requirements regarding data sovereignty or geolocation.

The open standard also favors the interoperability between solutions from different providersA Spanish startup developing a SaaS product for, for example, document management in law firms, could showcase its capabilities as skills compatible with Claude, but also with other agents that adopt the same specification, thus expanding its market without having to redo integrations for each platform.

Furthermore, the partner ecosystem—with tools like Atlassian, Figma, Stripe, and Zapier—offers startups a shortcut: instead of building complex connectors for each service, they can leverage existing skills and focus on add layers of logic and personal experience on topThis fits well with the reality of many European companies, which operate with small teams and seek to maximize the return on each development sprint.

For CTOs starting to design their agent strategy, the lesson is clear: treat skills as long-term assetsversioning, monitoring, and improving them with real data, and aligning them with the control and governance layer that the organization defines. In this way, when the ecosystem matures—and standards stabilize—the company will already have its own catalog of capabilities, ready to be integrated wherever most appropriate.

Anthropic's opening of Agent Skills is redefining how AI agents are conceived in the enterprise: from general assistants controlled by prompts, to modular, portable, and auditable skills-based work platformsFor Spain and Europe, where regulatory pressure and the need for interoperability are especially high, this model offers an intermediate path between rapid innovation and rigorous control, leaving the door open for the true differentiating value to lie in the skills that each organization is able to build and govern.