You say into your phone, 'Book me a flight to Melbourne next Tuesday morning.' Within seconds, your AI assistant has contacted a booking agent from a corporate travel supplier, checked flights and hotels—all in compliance with the company's travel policy. Throughout the process, you haven't sent a single email or made a single phone call. This is what Agent-to-Agent—A2A—is doing.

It's not just about letting two AIs chat; it's about enabling tasks to flow controllably between agents. In other words, you don't need to worry about how many steps were taken or how many systems were called—you just tell your agent what result you want. Your agent will find other agents, assess whether and how they can handle the task, delegate it, and then bring back the results.

01. From 'One AI' to 'The Internet of AIs'

When Google launched the A2A protocol a little over a year ago, many people were unclear about what it meant. At the time, the debate centered on which large model was smarter, akin to arguing over which computer had a faster CPU—without considering that if these computers couldn't connect to the internet, they'd remain isolated islands.

That's no longer the case. By 2026, the AI protocol stack has formed a clear hierarchy: MCP handles 'micro-execution'—how agents call tools; A2A manages 'macro-collaboration'—how agents interact. According to media citing a Gartner report, the widespread adoption of MCP and A2A marks the official end of the 'isolated system' era for AI Agents, ushering in a 'TCP/IP moment' for intelligent agents. Previously, every AI application reinvented the wheel, writing custom API adapters; now, based on a unified protocol, agents can collaborate seamlessly across applications, organizations, and clouds, much like humans use USB-C interfaces and HTTP protocols.

This protocol is now maintained by the Linux Foundation and supported by over 150 organizations, including Microsoft, AWS, Salesforce, and SAP. Industry partners such as China Mobile, Orange, and AIS have jointly released the A2A-T protocol for the telecommunications sector. In June this year, the State Administration for Market Regulation also released China's first national standard system for intelligent agent interconnection, covering seven core areas: overall architecture, identity codes, agent discovery, and agent interaction. Rules for how agents 'shake hands,' 'communicate,' and 'trust each other' are being written into standards.

02. The Numbers Speak for Themselves

How fast is this trend? Gartner's 2026 CIO survey shows that only 17% of organizations currently deploy AI Agents, but over 60% expect to do so within the next two years—the most aggressive adoption curve among all emerging technologies. By the end of 2026, 40% of enterprise applications are expected to integrate task-oriented AI Agents, up from less than 5% in 2025.

According to the CIDC's '2026 China Enterprise-Grade AI Agent Industry White Paper' (released simultaneously by CCTV), the market size for enterprise-grade AI agents in China reached 21.2 billion yuan in 2025 and is expected to grow to 44.9 billion yuan in 2026, with the potential to exceed 332 billion yuan by 2029. Meanwhile, IDC predicts that the global AI Agent market will surpass 1.2 trillion yuan by 2026.

McKinsey's research shows that over 38% of medium-to-large enterprises globally have deployed autonomous decision-making agents in at least one business line, up from less than 12% last year. Nearly 60% of operational costs for Agentic AI are spent on verifying and optimizing responses, rather than generating initial outputs—indicating that the real bottleneck is no longer 'can AI think?' but 'can AIs collaborate effectively?'

03. Agents Are 'Connecting' in the Business World

Behind these numbers are real-world scenarios.

Cross-border e-commerce is one of the earliest fields to adopt the A2A model. According to industry media, at the Jakarta International Expo Center, an overseas buyer submits a procurement request on their phone. The buyer's assistant agent automatically reads the category and specifications, filters potential suppliers from thousands of booths, and sends inquiries to the seller's end. The seller's digital avatar agent follows up with questions about quantity, budget, and delivery requirements while retrieving product information from the enterprise knowledge base. A backend matching agent further verifies factory capacity and inquiry value. By the time the buyer and seller meet in person, the most time-consuming screening and clarification work has already been completed by the agents. The '2026 China Export Cross-Border E-Commerce Development Trends White Paper' reveals that over 98% of surveyed Chinese sellers use AI tools when operating Amazon stores, with 16% advancing from standalone tools to deploying AI workflows or agents. On the 1688 platform, 20% of procurement requests are now initiated by AI agents.

The finance and taxation industry is undergoing a similar transformation. A fintech company's AI accountant handles bookkeeping, while its AI invoicing agent issues invoices. However, a standalone AI accountant can only follow generic rules, whereas each industry has its own accounting language. The real value lies not in replacing an accountant with an AI but in enabling the AI accountant and AI advisor to collaborate like seasoned and novice accountants—one handling bookkeeping tasks, the other injecting professional knowledge based on industry knowledge graphs. The A2A research report released by Tsinghua University's Shenyang team points out that future enterprise competition will not just be about model capabilities or software features but about designing effective agent collaboration networks.

Even everyday smartphone operations are being transformed by A2A. Honor's YOYO and WeChat achieved the first A2A collaboration—users dictate information into their phones, YOYO identifies the intent, and delegates the task to WeChat, which precisely pushes it to the designated recipient. The user doesn't need to open WeChat or type manually; key steps are completed after user confirmation.

04. When Agents Make Decisions for You, Who Watches Them?

However, letting agents freely 'connect' also introduces new challenges.

A2A describes interactions between two or more AI agents—they may come from different organizations or the same one. The most complex and valuable scenarios involve cross-organization collaboration: your personal agent needs to interact with external merchant agents. This raises governance issues: Who is responsible if an agent autonomously makes a purchase beyond budget? How is accountability traced if two agents make an error during negotiation? The Tsinghua research report proactively identifies these issues: responsibility delineation due to algorithmic black boxes, security risks from malicious interactions between agents, and the lack of standards for cross-platform interoperability.

Gartner's 2026 Hype Cycle for Agentic AI specifically notes that governance, security, and cost management have become as important as core technologies. Security vendors like Okta have launched solutions tailored for agent-to-agent connections, ensuring every task handoff is traceable and auditable. While Alipay's AI payments have completed 300 million intelligent agent transactions, most scenarios still require user authorization, falling short of true 'agents paying on your behalf.' Juniper Research predicts that global autonomous agent commerce will total $8 billion in 2026, growing to $3.5 trillion by 2031—a figure that underscores both potential and current restraint.

05. Conclusion

We are undergoing a shift from 'teaching AI to think' to 'teaching AI to socialize.' In 2025, the focus was on model parameters and reasoning power; by 2026, it's on how well agents can find each other, collaborate reliably, and deliver results securely.

A2A is transforming every AI agent from a standalone tool into a member of a collaborative intelligent agent network. In the future, you may not care which model or platform you use—you'll just know that your agent can find the right agent to get things done.

- End -