Author: Xiaojing, Tencent Technology

In 2011, Marc Andreessen wrote "Software is eating the world." In 2026, Fortune used one sentence to summarize the current situation: "The thing that eats the world is being eaten."

In April 2026, Notion's product team wrote this in the official blog: We have received a large number of requests from enterprise customers. They want to use Notion in AI-first workflows, and they also want to directly access Notion's workspace from tools like Cursor and Claude. So they released the Notion official MCP Server.

Adobe announced the launch of CX Enterprise at the Summit conference on April 20, 2026. This is a new architecture that repackages the capabilities of all its products into Agent Skills and MCP endpoints. The list of partners includes Anthropic, Google Cloud, Microsoft, OpenAI and AWS. Adobe's own creative and marketing workflows will now exist in the form of callable capabilities.

These companies are vying to turn themselves into capabilities callable by Agents. What are the software companies vying to be “swallowed by AI” thinking about?

In the past, the distribution logic of software was "users open the application", and users took the initiative to find tools. Now, a new distribution logic is taking shape: when the Agent handles tasks on behalf of the user, it will decide which external capabilities to call during task execution. Notion does not actively enter this logic, and its workspace is invisible to the Agent.

The software industry was “collectively abandoned” by capital

In February 2026, when Anthropic released Claude Cowork and OpenAI released Codex for (almost) everything, software stocks fell sharply. Morgan Stanley analyst Keith Weiss called it "SaaSpocalypse". Thomson Reuters plummeted about 16% in a single day, LexisNexis parent company RELX's London stock price fell 14%, and the iShares technology software ETF has fallen more than 28% from its peak in September 2025.

The logic of the market is: If AI Agent can directly operate any software with an interface, there is no need to pay a seat fee for each SaaS tool. IDC predicts that by 2028, the pure seat pricing model will be eliminated, and 70% of software vendors will have to reconstruct their pricing strategies and shift to new value indicators such as consumption, results, or organizational capabilities.

But this "SaaS is dead" narrative, Morgan Stanley believes is an "overreaction." There is a sentence worth noting in their research report: In this round of changes, the "strongest athletes" in the best position are Microsoft, Salesforce and ServiceNow - they happen to be companies that do not wait for AI to replace themselves, but actively connect themselves to the Agent ecosystem.

Klarna CEO Siemiatkowski himself later clarified this judgment: "I don't think this is the end of Salesforce, it may even be the opposite. What is more likely to happen is that fewer SaaS companies will integrate the market and provide what they do to other companies." This shows that the way out for software companies is not to "be replaced by AI", but to "make themselves capable of being called by AI."

Skill and Plugin: one is content and the other is container

In this transformation, Skill and Plugin are the two most obvious forms, and they are also the most commonly used packaging methods when software companies actively access the AI ecosystem.

According to the developer documentation of OpenAI Codex and Anthropic Claude Code, the division of labor between the two is as follows:

Skill is the encoding format of behavior. Its physical form is a directory containing the SKILL.md file. Metadata is declared in YAML. The text is an execution instruction written in natural language. It is used to tell the Agent what steps to follow when encountering certain types of tasks, what rules to follow, which tools to call, and what results to output.

Agent will scan the metadata summary of all Skills before processing the task, and only load the complete content when it decides to call a Skill to control context consumption. The scope of a skill is a single warehouse or local workspace. There is no version number, no permission declaration, and no mechanism for cross-team distribution.

Plugin is a container for encapsulation and distribution. The original words of the OpenAI Codex document are: "For reusable distribution of your own skills, prefer plugins." Plugin is packaged as a complete unit of capabilities: Skills (workflow instructions) + App integrations (external application integration) + MCP server configuration (external system connection).

Anthropic's Claude Code Plugin also packages Hooks (control nodes for executing the life cycle), Agents (sub-Agent definitions), and Commands (slash commands) in addition to Skills. Plugin has a version number, manifest, permission statement, privacyPolicyURL, termsOfServiceURL, brand color, and screenshot fields. The existence of these fields itself shows that the Plugin is designed for an open commercial directory and has a structure that can be reviewed, distributed, and priced for sale.

The relationship between the two is clearly stated in the documents of both companies: Skill is a creative medium for local iteration, and Plugin is a circulation commodity for distribution. If a software company only releases an MCP Server, it will solve the problem of "how does the agent connect to me"; if it further packages workflow knowledge, execution constraints and permission configuration into a plug-in, it will solve the problem of "how does the agent make good use of me". The latter has a deeper moat and is closer to real commercial assets.

MCP is a connection pipeline, which addresses the accessibility of data and operations; Skill and Plugin are behavioral containers, which address the knowledge encapsulation of how tasks are completed. The former is infrastructure, and the latter is the application layer that truly embodies professional value. If we want to talk about potential for commercial value, Skills and Plugins are more like apps in the AI ​​era.

Software that proactively seeks change

In the face of this change, software companies have taken the initiative to become the "Skill" or "Plugin" of large models or to "connect large models" with MCP, but the focus behind them is different.

Microsoft and Salesforce have taken the most radical actions, with very similar goals: to turn themselves into the base of the Agent workflow and let the skills of other tools run on their own. Microsoft's Declarative Agents framework allows companies to build exclusive Agents directly in Word, Teams, and Outlook. In April 2026, MCP Apps will further allow MCP Apps to render interactive UIs in the Copilot dialog window. Users can approve expenses, check project progress, and compare documents without leaving the chat interface. monday.com is one of the first partners.

The CRM data of enterprise customers is in Salesforce. Agents cannot bypass this data if they want to handle sales and customer service tasks, and they cannot bypass the orchestration layer of Salesforce. AgentExchange lists more than 1,000 skills and agents from more than 200 partners, but what really locks customers in is the business data stored in Salesforce.

Agentforce 2.0 directly charges Sales Development and Sales Coaching skills at US$2/conversation, which is currently the clearest case of direct pricing for enterprise-level skills. Marc Benioff said that enterprises must become "Agentic Enterprises", apparently "persuading" users that the pivot of this transformation should be completed on Salesforce.

ServiceNow has never sold a Skill separately, but its logic is no different from Microsoft and Salesforce. Whoever's IT processes, HR processes, and financial approvals run on ServiceNow, Agents collaborate in the ServiceNow environment, and the value of workflow capabilities is naturally realized through subscription premiums. ServiceNow's full-year subscription revenue will exceed $13 billion in 2025, a year-on-year increase of approximately 22%. CEO Bill McDermott said on the earnings call that AI will create approximately $500 million in value within the company in 2025, and that IT service desk first-level work orders can be processed 99% faster than humans.

Adobe's transformation efforts are the largest among traditional software companies. As mentioned at the beginning of the article, at the Summit conference in April, Adobe restructured its entire Experience Cloud into a system with Agent Skills and MCP endpoints as the core, covering site optimization, data insights, audience creation, and content supply chain. It provides three deployment modes and supports direct calls in Anthropic, Google Cloud, Microsoft, and OpenAI environments. Behind this architecture is a set of figures from Adobe's own research: 75% of companies list data integration as the primary obstacle to AI implementation, 71% face talent gaps, and 68% believe that ROI is unclear. Adobe encapsulates its marketing and creative workflows into callable Skills, betting that companies are willing to pay for ready-made professional workflows in the face of these three obstacles.

Notion, Stripe, and Shopify are more like being pushed. Notion wrote in his blog that the official MCP Server was released because enterprise customers required direct access to the workspace from Cursor and Claude. Stripe runs remote endpoints at mcp.stripe.com, and Shopify has launched four official MCP Servers, which only allow Agents to call themselves when processing payment and e-commerce tasks. This type of company's "access to the big model" is more like a defense, which does not generate new revenue, but can prevent it from disappearing from the user's workflow.

Docusign goes a step further. It packages the signing process into a Skill and puts it on AgentExchange instead of just publishing an MCP Server. The gap between the two is worth clarifying: MCP endpoint solves "Agent can connect me", and Skill encapsulation solves "Agent knows how to use me well". Docusign’s data is that after Skill was launched on the shelves, more than 200 private quotes were processed in Q4 of 2025, and the signing time was accelerated by 60%. From access to encapsulation, it is a step from infrastructure to application layer. The commercial value gap of this step is obvious.

The access methods of domestic companies basically stay at the MCP layer, solving the problem of "Agent can connect me". Not many companies have further encapsulated workflow knowledge into Skills like Docusign. For example, Alipay's MCP Server solves the technical connection of payment, but the workflow knowledge of "how agents can optimally use Alipay in different business scenarios" has not yet been systematically encapsulated.

Feishu has launched MCP capabilities for document scenarios on the open platform, DingTalk’s official team has open sourced dingtalk-openclaw-connector, and Amap has encapsulated navigation data into MCP Server and put it on the shelves. MCP Plaza currently contains more than 9,000 services.

But at present, few companies have systematically encapsulated the workflow knowledge of "how the agent can make good use of me in different business scenarios" into Skills.

"There is no clearly leading model party in China such as OpenAI or Anthropic to take the lead in breaking out of some industry paradigms. Big manufacturers have their own ecology, but they do not have particularly leading or outstanding features in terms of model capabilities. This may also be the reason why application layer companies are slow to follow up." A primary market investor in the technology field told Tencent Technology.

However, some business leaders of mobile phone terminal manufacturers said, "In the future, Agent will be the OS and Skill will be the mobile APP. We are also actively thinking about how to promote it."

"Skill currently lacks unified standards and has low production thresholds, so it is difficult to say that it has commercial value; in contrast, Plugin is more like a commercial form. However, if the model capabilities are further improved, will the model swallow everything up? This is also a current discussion among everyone." A software Technical experts in the software industry said, "Currently, it seems that everyone in the domestic application layer and software industry is still running through the connection layer. This is related to the maturity of the overall domestic Agent ecosystem, and also related to the huge gap between the prosperity of the domestic SaaS ecosystem and the intensity of competition compared with foreign countries."

What skills really have commercial value?

So, does the popular Skill have commercial value? Skill, Plugin, or software’s built-in Agentic AI capabilities, what is the greatest possibility to break through the business closed loop? The jury is still out.

Gartner's forecast provides a macro coordinate: by 2028, 33% of enterprise software will contain Agentic AI capabilities, and in 2024, this number will be less than 1%. However, the same research also shows that more than 40% of Agentic AI projects will be stopped before the end of 2027 due to out-of-control costs, unclear value, or lack of risk management and control. There will be a high concentration of winners.

Skills that can generate real business value have three things in common:

First, it carries execution knowledge in a specific field, not just a general process description. Harvey is the clearest case. The legal AI company calls on models from Anthropic and OpenAI under the hood, but its core value lies in the layers of expertise it builds on top of its models: content partnerships with LexisNexis, a data privacy architecture for law firms, industry methodologies gained through hiring former BigLaw attorneys. Harvey's system processes more than 400,000 Agentic queries every day and has more than 25,000 user-created custom workflows. In March 2026, Harvey completed US$200 million in financing at a valuation of US$11 billion. CEO Winston Weinberg said: "Harvey is not just Claude with legal reminders added. It is a platform built for the way legal teams actually work, including expertise, workflow integration, compliance infrastructure and institutional knowledge. That is the focus."

Second, include verifiable quality standards and failure handling logic. The Skill provided by clearMDM on Salesforce AgentExchange is used for Salesforce CRM data quality control. The starting price is about 100 pounds/month/company. The core is a set of workflow logic that specifically handles data cleaning scenarios, including how it handles various abnormal inputs. What companies buy is stability and predictability, not clever reminders.

Third, it has auditable execution constraints. The stronger the model, the more the company needs to limit its autonomy. A Skill that can truly be implemented in an enterprise must be able to clearly explain what operations it will perform under what circumstances, and at which nodes it will request manual confirmation. Adobe clearly provides two different levels of manual supervision mechanisms in CX Enterprise. This design reflects the real needs of enterprise customers.

In contrast, skills such as general writing, ordinary meeting minutes, and simple copywriting have been absorbed by the model's native capabilities and have no independent commercial value. Tool-enhanced skills, such as PPT generation, Excel analysis, and code refactoring, are more likely to be directly built into products such as Office, Notion, and GitHub Copilot, and exist as functions rather than skills.

What may the future of the software ecosystem look like

After this, there is a bigger question: What will the structure of the software ecosystem become?

Morgan Stanley analyst Keith Weiss summarized the impact of AI on the software industry as the "trinity of software fears" - seat pricing is threatened, the logic of self-built alternatives to outsourcing is reversed, and AI directly enters the application layer competition. But Morgan Stanley also believes that this concern "gives too little credit to software giants for their ability to participate in this round of innovation" and lists Microsoft, Salesforce, and ServiceNow as the companies most likely to win.

IDC's judgment on this change is that the next generation of enterprise technology stacks will be built around AI Agents rather than SaaS interfaces. "Today's technology stack is built around SaaS interfaces, and tomorrow's technology stack will be built around AI Agents that interact with modular back-end services." Data lakes and real-time data connections have become key enabling factors, and supplier relationships have evolved from "UI-centric participation to 'Agent-centric enabling cooperation.'"

This judgment means that the three types of software will have different directions:

Software with network effects and proprietary data will be stronger. Airbnb's host network, Uber's driver network, and Palantir's defense data cannot be replaced by agents. Instead, AI will enhance their ability to process these data and networks. The response strategy of such companies is to proactively encapsulate their data and core logic into high-value skills that can be called by the Agent, rather than waiting for the Agent to bypass them.

SaaS, with interface as its core moat, faces real threats. If the user value mainly comes from a beautiful UI and click process, but there is no unique data and workflow accumulation behind it, then the Agent can either control the interface or directly bypass it and call the API. The way out for this type of product is to reconstruct itself into an Agent orchestration layer, with permission management, process verification and audit capabilities as its core.

Industry vertical software is the clearest opportunity for skill commercialization. Harvey's path in the legal industry and Rogo's path in the investment research industry prove one thing: when you deeply encode the knowledge of executing real tasks in an industry into Plugin and reach large customers through corporate procurement channels, this path does exist, and valuation data has proven its commercial scale.

Finally, going back to the batch of active skills, the most important thing may not be how much money can be sold in the "Skill Market" and "Plugin Market". What affects life and death is whether these capabilities can be called upon at the moment when the task occurs when the Agent handles the work on behalf of the employee.

This is a fundamental reconstruction of the software entrance. The entrance has changed from a desktop icon to a list of tools called by the Agent when performing tasks. Software that does not appear on this list will be silently bypassed without the user even realizing it.