Is the era of Agents here, and will GUI no longer monopolize the software entry point?

Since the beginning of 2026, a change in the Internet industry that was originally hidden behind the interface has been slowly surfacing.

Google, Atlassian, and Google Cloud have successively launched the Developer Knowledge API, Rovo MCP Server, and fully - managed remote MCP services. On the surface, these actions are scattered across documents, collaboration software, and cloud platforms, but they all point to the same thing: as Agents enter the workflow, software companies are building a new layer of entry points for Agents.

An Agent is an intelligent entity that can call tools and execute tasks on its own; an API is an external call interface for software; MCP, short for Model Context Protocol, is a set of general protocols that connect large models with external tools, data, and systems; CLI, or command - line interface, is an interaction method more suitable for programs and Agents to directly invoke capabilities; skills can be understood as a set of packaged tool capabilities that let Agents know what they can do and how to call them.

Together, they point to the same thing: making it easier for Agents to find, connect to, and use tools.

Looking back, although Google, Atlassian, and Google Cloud have different actions, their directions are very similar. They are all organizing the capabilities originally hidden in documents, workflows, cloud resources, and product features into new entry points that Agents can directly discover, understand, and call.

In the past, people mainly used the GUI, or graphical interface, to complete operations step by step through pages, buttons, menus, and forms. Now, software is growing another layer of entry points for Agents to directly invoke capabilities.

The GUI is still there, but it no longer monopolizes the software entry.

This is not far from ordinary people. Today's AI mostly only "talks". To move towards "doing", it needs to be able to call documents, calendars, applications, and processes, complete tasks on its own, and directly provide results to users.

01

Software companies are building a new layer of entry points

The recent wave of changes, on the surface, appears as scattered product actions: some release CLIs, some launch MCP servers, some turn documents into machine - readable interfaces, and some reorganize APIs, plugins, and knowledge bases originally for developers into a tool layer that Agents can call.

Although the actions are scattered, the directions are consistent. Software companies are all dealing with the same problem: after Agents enter the workflow, how can software capabilities be more stably discovered, understood, called, and executed.

In recent months, a series of actions by Google, Atlassian, and Google Cloud have made the outline of this change clearer.

In February 2026, Google released the Developer Knowledge API and a supporting MCP server. On the surface, this seems to be about processing development documents, but it actually points deeper. Google aims to organize the knowledge originally scattered across web pages, instruction manuals, and tutorial pages into an official source that programs can call, models can read, and Agents can access. Development documents are not just auxiliary materials; they are the instruction manuals and usage boundaries of product capabilities. By allowing machines to read this content, Google shows that in the future, reading documents, finding capabilities, understanding rules, and calling interfaces will not be done only by humans.

Atlassian's actions have brought the same change to enterprise collaboration software. In February 2026, Atlassian announced the general availability of the Rovo MCP Server, allowing external AI clients to access the workflows and knowledge content in Jira and Confluence. The official statements repeatedly emphasize security, permissions, and governance, which is crucial. It shows that this kind of access is no longer just a technical demonstration but is gradually entering real organizational environments. Originally, people had to enter the interface layer by layer to access tasks, documents, and knowledge. Now, this content is open to Agents through another channel.

Google Cloud's progress has extended this from a single product to the platform level. In December 2025, Google Cloud announced official MCP support for Google services and Google Cloud services and launched fully - managed remote MCP servers. The significance of this action is not just "supporting MCP" but also reorganizing the originally complex cloud service capabilities into a framework more suitable for Agents to call. Cloud platforms already have many resources, interfaces, and permission controls. What Google Cloud is doing now is to allow Agents to access these capabilities more directly without having to explore through layers of documents and consoles.

Looking at these companies together, the outline is clear. Although they have different products, scenarios, and target audiences, they are all doing the same thing at the core: extracting software capabilities and creating entry points that machines can understand more easily and Agents can call more readily.

This can be understood as the "double - layer expansion" of software entry points. One layer continues to face humans, responsible for display, browsing, configuration, collaboration, and confirmation, which is still the strength of the GUI. The other layer mainly faces Agents, responsible for capability discovery, parameter transfer, task execution, result return, cross - system call, and automated orchestration.

Google turning document knowledge into a machine - readable entry point, Atlassian connecting enterprise workflows and knowledge content to external AI clients, and Google Cloud organizing service capabilities into fully - managed MCP services together show that this is not a sporadic attempt but an industry - wide adjustment.

From this perspective, software companies' understanding of "entry points" is no longer limited to a screen, a homepage, a console, or a navigation bar. An entry point can also be a command, a machine - readable document, an MCP server, or a skill description that Agents can recognize and call. The connection between humans and software has not disappeared; it has just added another layer of channels.

02

When the execution subject expands from humans to AI

At the AI DingTalk 2.0 launch event on March 17th, Chen Hang, the founder and CEO of DingTalk, mentioned a judgment: The real big change in the AI era is not just that the models have become more powerful, but that the main body of the Internet is expanding from "humans" to "AI".

This statement can explain why CLI, MCP, and A2A have become popular recently. The key lies not in the sudden popularity of new terms but in the change of the software users. Humans are still there, and AI has also entered. Once software serves both types of subjects, many problems originally hidden deep in the system will emerge.

The reason why the GUI has long been the main form of software is that it is most suitable for humans. Humans can look at pages, find buttons, and complete operations step by step along navigation, forms, and processes. Many software capabilities were originally organized in this way.

However, AI does not work like this. AI does not need to understand the page first and then explore the path step by step. It is more concerned about what capabilities are in the system, how to call the interfaces, how to transfer parameters, how to return results, where the permission boundaries are, and which actions can be executed next. At this point, the capabilities originally wrapped in the interface need to be re - organized. The pages are still there, but the pages themselves are no longer enough.

So, although it seems that terms like CLI and MCP are becoming popular on the surface, the real change is in the organization of software.

In the past, people cared more about "how to click", and now they care more about "how to call". The former is the interface logic, and the latter is the capability logic. The interface logic mainly serves humans, while the capability logic increasingly serves AI.

If a software only has buttons, menus, and page paths, it is difficult for AI to use it efficiently. However, once these capabilities are organized into clearer interfaces and more direct call methods, the relationship between software and AI will be completely different.

The change at the protocol level makes this clearer. In December 2025, Anthropic donated MCP to the Agentic AI Foundation under the Linux Foundation. The significance of this action is that how AI connects to external tools, data, and systems is no longer just an internal optimization of a single company but has become a common foundation for the entire industry.

Because once AI truly enters the workflow, it cannot stay in the chat box forever. It must connect to documents, databases, enterprise applications, and cloud resources. If the connection methods are fragmented in private interfaces of different companies, the cost will be high, and the ecosystem will be scattered. Putting MCP in a more neutral framework is essentially building a foundation for "how AI connects to the world".

Google's promotion of A2A is in the same direction. A2A focuses on how AIs exchange information and coordinate actions with each other, rather than how an AI calls a single tool. That is to say, in the future, not only will an AI call tools, but an AI will also connect to another AI and string together capabilities in multiple systems.

At this point, the connection relationships in the software world have changed. It is no longer just "humans open pages and complete tasks along the path", but will increasingly become "AI calls tools, AI finds other AIs, and cross - system collaboration".

A recent engineering detail disclosed by Cloudflare also illustrates the problem. It mentioned that if its huge API system is directly exposed to AI, the cost will be very high. An equivalent native MCP server would consume 1.17 million tokens, exceeding the context window of many models. To solve this problem, it proposed Code Mode, which compresses large - scale API capabilities into a more compact form, allowing AI to access "the entire API" with two tools and about 1000 tokens of context.

This example shows that in the AI era, it is not enough to just open up capabilities. Software also needs to reorganize, describe, compress, and manage capabilities. Many software already had APIs, but that does not mean they are naturally suitable for AI. Whether AI can discover capabilities at low cost, accurately understand capabilities, and stably call capabilities has become a new product engineering task.

Therefore, the challenges faced by software companies have also changed. After software serves both humans and AI, it is not just about having a user - friendly interface. It also depends on who can organize capabilities more quickly, connect processes, and manage risks. In the end, this is not just about adding a few new features; it means that software needs to be rebuilt.

This change will also be reflected in budget reallocation.

In the past, software companies' investments mainly focused on interfaces, interactions, paths, and growth, aiming to make it easier for humans to understand, operate, and stay. After Agents enter the workflow, more money will flow to another area: capability reconstruction, protocol access, governance enhancement, and optimization of inference and call costs.

Cloudflare's efforts to reduce the token overhead when exposing APIs to AI and Atlassian's strengthening of permission control, usage logs, and governance capabilities in the Rovo MCP Server indicate that in the future, software companies will compete not only in interface experience but also in making capabilities more accessible, easier to manage, and more cost - effective.

03

The focus of competition is shifting behind the interface

The investment circle has conducted in - depth research on this change. On March 18th, Zhong Tianjie, an investment director at ZhenFund, proposed in a signed article that after the arrival of the Agent era, the focus of competition among software companies is shifting from interface experience to callable capabilities. He even further judged that "we may no longer invest in software companies with a GUI - centric mindset".

This article is not the final conclusion of the industry, but it has put a growing issue on the table: when AI becomes the new execution subject, the product logic, competition logic, and value logic of software companies built around the GUI in the past need to be re - evaluated.

Beyond the investment circle, the official statements of large companies are also changing. The latest financial reports show that although neither Alibaba nor Tencent has directly included CLI in their performance narratives, the directions of "AI entering real - task execution" and "AI - native applications entering the growth logic" are becoming clearer.

Alibaba mentioned that Qwen has been integrated into scenarios such as Taobao, instant retail, Gaode, Fliggy, and Alipay, and clearly stated that AI agents are executing real - world tasks on a large scale. Tencent has focused on AI improving productivity, driving R & D and infrastructure investment, and the promotion of business ecosystems by AI plugins and AI - native mini - programs.

Terms like CLI and MCP have not officially entered the financial report language, but "AI moving from chatting to execution" and "AI entering real workflows" are getting closer to the official statements of large companies.

Looking at the actions at the domestic product level, the direction is also very clear. Since 2026, several mainstream software and cloud platform companies in China have successively promoted MCP, Agent plugins, tool - based capabilities, and CLI - based transformations to the forefront. Although the actions are scattered, they all point to the same thing: organizing the capabilities originally hidden behind the interface into tools, services, and interfaces that AI can call.

At this point, there are roughly three layers that domestic software companies really need to improve.

The first layer is to organize capabilities. Many software products have accumulated a large number of functions over the years, but these functions have long been organized based on "how humans use them" and are hidden behind pages, buttons, menus, and forms. Humans can click through layers to find them, but AI cannot. AI needs clear tool definitions, clear action boundaries, and stable return results. The recent actions of Feishu and Alibaba Cloud are most notable in this regard. What they are doing is not just adding a new interface but reorganizing existing capabilities into a layer of tools more suitable for AI to call.

The second layer is to integrate these capabilities into more general protocols and ecosystems. Even if the capabilities are extracted, if they are connected and described differently by different companies, this new layer of entry points will be difficult to develop. Tencent Cloud has incorporated MCP into its plugin system, and Alibaba Cloud has integrated official MCP services and custom MCP deployments into its platform. Behind these actions is the answer to the same question: In the future, AI will not stay in a single product of a single company but will shuttle between more workflows, platforms, and toolchains. The company that integrates its capabilities into the mainstream protocol first will have a better chance of being integrated into the next - generation Agent workflow.

The third layer is to improve governance. This step is more difficult and more practical. Whether AI can perform tasks is important, but in the enterprise scenario, what really determines whether a company dares to put AI into real processes is another set of issues: who grants permissions, whether calls are recorded, whether errors can be rolled back, how to handle misoperations, and how to isolate sensitive data. What is notable about DingTalk's "Wukong" release is not only the CLI - based transformation and atomicization of capabilities but also the repeated emphasis on permission control, security sandboxes, snapshots, and roll - back systems. Because once AI touches approvals, documents, spreadsheets, knowledge bases, and business systems, the risks will also increase. Enterprises will not only ask "how powerful it is" but also "what to do if it goes wrong".

Looking at it this way, what software companies really need to do next is to promote three things simultaneously: first, organize capabilities from behind the interface; second, integrate capabilities into more general protocols; and finally, improve permissions, auditing, roll - back, and security boundaries. Missing any step will not work. Without a protocol, the capabilities will be scattered; without governance, enterprises will not dare to put AI into core processes; without capability reconstruction, AI can only stay at the surface - level interaction and cannot enter the deep system.

This will also change the position of the GUI. The GUI will not disappear. Browsing information, initiating collaboration, making judgments, conducting reviews, and making final confirmations often still rely on the interface. The change is that it no longer monopolizes the software usage path. A mature software system is increasingly like a two - layer structure: the upper layer is the interface for humans, responsible for display, collaboration, and confirmation; the lower layer is the capability network for AI to call, responsible for execution, scheduling, and connection.

In the end, it is no longer a question of "whether to embrace AI". The more realistic question is whether software companies are willing to rebuild themselves. Extracting the capabilities originally in the GUI, re - defining the processes originally only open to humans, and turning the risks originally handled manually into system - level constraints. The company that completes this step first will have a better chance of occupying a position in the next - round of entry - point changes.

Because once the AI entry point is truly established, software competition will not only be about usability but also about who has a more callable, connectable, and workflow - integrable capability system.

This article is from the WeChat official account "XiuTai", author: Shi Can, published by 36Kr with authorization.