I Think 2027 Will Be the Breakout Year for A2A
As personal agents become a new gateway to the internet, businesses will need agents of their own. Negotiating goals, delegating tasks, and making verifiable commitments could drive the next agent ecosystem.
As personal agents become a new gateway to the internet, businesses will need agents of their own.
I think 2027 could be the year when collaboration between agents starts entering commercial use at scale.
Not because a communication protocol suddenly becomes mature, but because the internet is beginning to have agents that can represent different parties over time. Those agents need a new way to work together.
Here, I use A2A—Agent-to-Agent—to describe a form of interaction: agents representing different parties negotiate and collaborate around goals, constraints, proposals, and commitments. I am not referring exclusively to any particular protocol.
My judgment comes mainly from two changes already underway. Personal agents are becoming a new entry point for users, while businesses cannot simply give up their knowledge and decision-making authority. The meeting of these two forces could mark the beginning of the next agent ecosystem.
Why hasn’t A2A really taken off before?
On April 9, 2025, Google introduced the Agent2Agent protocol to address communication and collaboration between agents built by different vendors and with different frameworks. Yet for a considerable time afterward, at least from what I observed, we did not see large numbers of ordinary users completing real tasks through collaboration between different agents.
I think the fundamental reason was not poor protocol design. The underlying conditions were not ready. Most agents could not reliably complete a complex task on their own. Tool use, environment management, long-running execution, and recovery remained unresolved. When one agent is unreliable, having several work together only amplifies the uncertainty.
More importantly, there was little real demand. Users were still experimenting with getting agents to help with their work. Few people actually had an agent that could act on their behalf over time. Businesses therefore had little incentive to build a dedicated entry point for agent collaboration.
Protocols can be standardized ahead of time. Real demand for collaboration cannot be created by a protocol.
Personal agents are becoming a new gateway to the internet
Watching coding agents develop over the past year, I have felt a clear shift in what we discuss.
Early on, the questions were whether an agent could write code or call tools. Now, many discussions focus on how to keep an agent working on longer tasks, manage its environment, recover from failures, and reduce its dependence on constant human intervention.
This does not mean agents are reliable enough to work entirely on their own. But in some fields, they have gradually moved from demonstration tools to executors that can take on real work.
Another important change is underway: agents are moving from tools that users actively launch toward personal representatives that can keep working.
Meta introduced Muse on September 8, 2026, and OpenAI introduced Dots on September 29. What makes these products worth watching is not just model capability, but their product form: their own cloud computing environments, the ability to remember user preferences, continued work on delegated tasks, and requests for confirmation when needed. These are their product directions and officially described capabilities; they do not establish that reliable ongoing execution has been broadly demonstrated.
With these capabilities, an agent becomes more than a client that calls tools. It starts to resemble a digital representative that can act continuously.
What could really change is user behavior.
Today, if I need a flight, I open an airline app. To arrange my schedule, I open a calendar. To buy something, I search an ecommerce platform. In the future, I may only need to tell my agent what I want to accomplish. It knows my preferences, has access to my schedule, understands my budget, and can keep making progress while I am offline. When a decision needs my input, it comes back to ask.
Once this becomes a habit, an important consequence follows:
A persistent layer representing the user’s interests starts to sit between the user and commercial services.
Businesses used to need to persuade users to open their apps. In the future, they may first need to persuade a user’s agent that their service is the right choice.
That could be a major shift in how internet traffic is directed.
Businesses won’t willingly surrender their business judgment
If personal agents become a new way for users to access the internet, how should businesses interact with them?
The most direct answer is to expose APIs: search products, check inventory, create orders, pay, and cancel. External agents can combine these calls. For many deterministic operations, that works perfectly well.
But a business’s real value usually lies not just in the functions it offers, but in knowing how to use those functions to solve a customer’s problem.
Consider a travel service. The valuable part is not simply retrieving flight information. It is knowing how to produce an acceptable proposal given a user’s preferences, budget, time constraints, and the service’s existing supplier relationships. That involves business experience, internal rules, resource relationships, and product design. For an established business, these are precisely the advantages it has accumulated over time.
When an external agent faces a complex request with an unclear goal that needs repeated clarification and adjustment, offering only a fixed set of functional interfaces often transfers much of the business understanding and solution design to that agent. Who decides which proposal best fits the user? Who explains the differences? Who has authority to adjust the terms or offer another option?
If the user’s agent does all of this, merchants may gradually lose the opportunity to define a solution together with the user. Over time, businesses could be reduced to interchangeable resource suppliers.
I do not think most competitive businesses will willingly accept that position. They are more likely to retain their business judgment and participate through a service representative that can understand external requests, propose solutions, and accept delegated work.
This does not require every company to run an independent large language model. A service agent can combine rules engines, workflows, APIs, and models. What matters is that it can represent the service in business interactions within the company’s authorized boundaries.
The personal agent represents the user’s interests. The service agent represents the company’s capabilities and commercial boundaries.
They do not need to share all their internal information or obey the same decision-making system. They need to find an outcome both can accept.
From calling functions to negotiating outcomes
Suppose I tell my agent:
Arrange a trip to Tokyo for under RMB 8,000. Prefer direct flights and a hotel with convenient transport. If prices are similar, I would like flexibility to change the itinerary.
With traditional system integration, my agent would call airline, hotel, and other service APIs separately, retrieve data, and then interpret business rules, compare prices, and develop a plan itself.
If the providers also have agents, things could work differently. The following is an imagined collaboration.
My agent could present the goals and constraints directly to an airline’s agent. Using its internal inventory, pricing policies, membership benefits, and cancellation and change rules, the airline’s agent could offer one or more proposals and explain their differences. If I care more about flexibility than the lowest price, my personal agent could continue the negotiation and ask whether a better combination exists. Hotels and other services could work similarly.
The personal agent ultimately chooses according to my interests. The provider’s agent offers proposals within its business capabilities and authority.
The difference from traditional API calls is not whether natural language is used. Structured messages could support this collaboration just as well. The real difference is:
We have mostly told software what operation to perform. In the future, we may start telling another party what outcome we want to reach.
The former centers on function calls. The latter centers on negotiating goals and delegating tasks.
Of course, when a task needs only a deterministic query or operation, an API remains the better choice. There is no reason to adopt A2A for its own sake. Service agents may add value over direct API calls when there are multiple viable solutions, incomplete information, constraints to clarify, or a need for ongoing collaboration.
I think this is where A2A will first establish commercial value.
Why do I think the turning point could come in 2027?
I do not think technological development follows an inevitable annual rhythm. Protocols appearing in 2025 and cloud agents appearing in 2026 do not prove that 2027 must be a breakout year.
But recent developments contain several signals worth taking seriously.
First, personal agent products are moving toward ongoing execution. Whether coding agents are taking on longer tasks or cloud-based personal representatives such as Muse and Dots are emerging, they are trying to move users from actively using AI each time to continuously delegating work to it. This shift determines whether real demand exists.
Second, commerce platforms have begun to take external agents seriously.
On January 11, 2026, Google and companies including Shopify introduced the Universal Commerce Protocol (UCP). It aims to standardize interactions between agents and commercial systems, from product discovery to purchase and post-purchase support. It supports communication through APIs, MCP, and A2A, while allowing merchants to retain their business requirements and transaction rules.
A more recent signal came on October 13, 2026. According to The Verge’s reporting, Shopify announced broader access to its product catalog and purchasing capabilities for third-party agents, extending beyond agents from specific partners. Users can configure their own agents and confirm purchases.
These developments mostly open product and transaction capabilities to external agents. They do not yet amount to full negotiation of goals across businesses. Their significance is that commerce platforms are beginning to treat external agents as a service entry point that deserves formal support, rather than merely a technical demonstration.
Internet products used to ask how to bring users into their applications. Some commerce platforms are now asking how to let external agents bring users’ needs into their transaction systems.
Once these two changes reach scale, they could reinforce each other. More personal agent users give merchants a stronger incentive to support this entry point. More services supporting agent collaboration let personal agents complete more tasks. Users then become more willing to delegate work to agents.
I think 2027 could be the year when this feedback loop starts accelerating.
That judgment has conditions. If, by 2027, personal agents have become a stable part of daily life rather than chat tools used occasionally; if more leading businesses formally support discovery, delegation, and transactions by external agents; and if tasks spanning services reach real business completion rather than merely demonstrate collaboration, I would consider A2A to be entering its breakout phase.
Conversely, if personal agents still lack real demand for continued use, or merchants offer only limited query interfaces without deeper task collaboration, my timing may be too optimistic.
I see 2027 as a window for a potential commercial turning point worth watching, rather than a certain technical milestone.
When agents can make commitments for others, the questions are only beginning
The hardest part of A2A may ultimately lie beyond communication. Two agents exchanging messages is different from agents completing a commercial transaction on behalf of different organizations.
If a personal agent claims it has the user’s authorization, how does a merchant verify that? If a provider’s agent gives a quote, how long is it valid? Is it binding? If a task fails, how do the parties establish responsibility?
When agents start making commitments on behalf of others, they need to exchange more than messages. They need verifiable authorization, commitments, and evidence of execution.
Google’s Agent Payments Protocol (AP2), introduced on September 16, 2025, has begun to address trust in agent payments through verifiable authorization credentials. I think infrastructure of this kind will become increasingly important. Collaboration between businesses ultimately has to produce real costs, returns, responsibilities, and fulfillment outcomes.
If one agent cannot prove what it is allowed to do, and another cannot prove what it has committed to, fluent communication alone will struggle to support a reliable business relationship.
That is why I think A2A’s development should reach beyond multi-agent frameworks and communication protocols. Identity, authorization, task state, transactions, billing, auditing, and dispute resolution may together form the infrastructure of the next generation of agent collaboration.
The next internet may connect representatives of different parties
Looking back at the internet’s development, we have become familiar with several forms of interaction. Websites give people access to online information. Apps put services in users’ hands. APIs let software systems exchange data and call capabilities directly.
Agents could introduce another shift. Users begin to have digital representatives that act for them over time. Businesses may have service representatives that express their capabilities, maintain commercial boundaries, and accept delegated work. These representatives could do more than exchange data: they could collaborate continuously around goals, constraints, quotes, commitments, and fulfillment outcomes.
What really changes is the relationship between decision-makers on the internet.
Most business judgment across applications has ultimately been exercised by people, with software carrying out the operations. In the future, some judgment may be exercised separately by agents authorized by different parties, who then collaborate to reach an outcome.
This does not mean the internet will automatically become more open. If personal agents become a new gateway for traffic, they could create new platform barriers. Whoever holds users’ long-term trust and delegation relationships may have an advantage in selecting services.
Whatever competitive structure emerges, businesses may have to reconsider one question:
When a user sends an agent to complete a task instead of visiting my product personally, how should I provide the service?
I think many internet products will face this question over the next few years. Competition may extend beyond getting users to open an app to getting their agents to choose a service. Reliability, product capabilities, past fulfillment performance, and support for agent collaboration could become new selection criteria.
That is why I think 2027 will be the breakout year for A2A.
Not because agents finally learn to talk to each other, or because one protocol is about to dominate everything. Rather, as more users have agents that can work for them over time, businesses will also need a way to represent themselves when interacting with those agents.
The internet has connected people, applications, and services. In the future, it may also connect agents acting for parties with different interests.
I find that more exciting than the A2A protocol itself.