Designing Collaborative Multi-Agent Techniques with the A2A Protocol – O’Reilly

It seems like each different AI announcement currently mentions “brokers.” And already, the AI group has 2025 pegged as “the yr of AI brokers,” typically with out rather more element than “They’ll be superb!” Typically forgotten on this hype are the basics. All people is dreaming of armies of brokers, reserving accommodations and flights, researching complicated matters, and writing PhD theses for us. And but we see little substance that addresses a essential engineering problem of those bold techniques: How do these unbiased brokers, constructed by totally different groups utilizing totally different tech, typically with utterly opaque internal workings, truly collaborate?

However enterprises aren’t typically fooled by these hype cycles and guarantees. As a substitute, they have an inclination to chop via the noise and ask the arduous questions: If each firm spins up its personal intelligent agent for accounting, one other for logistics, a 3rd for customer support, and you’ve got your individual private assistant agent attempting to wrangle all of them—how do they coordinate? How does the accounting agent securely go information to the logistics agent with no human manually copying information between dashboards? How does your assistant delegate reserving a flight without having to know the particular, proprietary, and sure undocumented internal workings of 1 explicit journey agent?

Proper now, the reply is usually “they don’t” or “with a complete lot of {custom}, brittle, painful integration code.” It’s changing into a digital Tower of Babel: Brokers get caught in their very own silos, unable to speak to one another. And with out that collaboration, they will’t ship on their promise of tackling complicated, real-world duties collectively.

The Agent2Agent (A2A) Protocol makes an attempt to deal with these urgent questions. Its objective is to supply that lacking frequent language, a algorithm for a way totally different brokers and AI techniques can work together without having to put open their inner secrets and techniques or get caught in custom-built, one-off integrations.

On this article, we’ll dive into the small print of A2A. We’ll have a look at:

• The core concepts behind it: What underlying rules is it constructed on?
• The way it truly works: What are the important thing mechanisms?
• The place it matches within the broader panorama, specifically, the way it compares to and probably enhances the Mannequin Context Protocol (MCP), which tackles the associated (however totally different) downside of brokers utilizing instruments.
• What we predict comes subsequent within the space of multi-agent system design.

A2A Protocol Overview

At its core, the A2A protocol is an effort to ascertain a method for AI brokers to speak and collaborate. Its purpose is to supply a normal framework permitting brokers to:

• Uncover capabilities: Determine different accessible brokers and perceive their features.
• Negotiate interplay: Decide the suitable modality for exchanging data for a particular job—easy textual content, structured kinds, even perhaps bidirectional multimedia streams.
• Collaborate securely: Execute duties cooperatively, passing directions and information reliably and safely.

However simply itemizing targets like “discovery” and “collaboration” on paper is simple. We’ve seen loads of bold tech requirements stumble as a result of they didn’t grapple with the messy realities early on (OSI network model, anybody?). Once we’re attempting to get numerous totally different techniques, constructed by totally different groups, to really cooperate with out creating chaos, we’d like greater than a wishlist. We’d like some agency guiding rules baked in from the beginning. These mirror the hard-won classes about what it takes to make complicated techniques truly work: How can we deal with and make trade-offs on the subject of safety, robustness, and sensible utilization?

With that in thoughts, A2A was constructed with these tenets:

• Easy: As a substitute of reinventing the wheel, A2A leverages well-established and broadly understood current requirements. This lowers the barrier to adoption and integration, permitting builders to construct upon acquainted applied sciences.
• Enterprise prepared: A2A contains sturdy mechanisms for authentication (verifying agent identities), safety (defending information in transit and at relaxation), privateness (guaranteeing delicate data is dealt with appropriately), tracing (logging interactions for auditability), and monitoring (observing the well being and efficiency of agent communications).
• Async first: A2A is designed with asynchronous communication as a main consideration, permitting duties to proceed over prolonged durations and seamlessly combine human-in-the-loop workflows.
• Modality agnostic: A2A helps interactions throughout varied modalities, together with textual content, bidirectional audio/video streams, interactive kinds, and even embedded iframes for richer consumer experiences. This flexibility permits brokers to speak and current data in essentially the most applicable format for the duty and consumer.
• Opaque execution: It is a cornerstone of A2A. Every agent taking part in a collaboration stays invisible to the others. They don’t must reveal their inner reasoning processes, their data illustration, reminiscence, or the particular instruments they is likely to be utilizing. Collaboration happens via well-defined interfaces and message exchanges, preserving the autonomy and mental property of every agent. Word that, whereas brokers function this manner by default (with out revealing their particular implementation, instruments, or mind-set), a person distant agent can select to selectively reveal facets of its state or reasoning course of through messages, particularly for UX functions, equivalent to offering consumer notifications to the caller agent. So long as the choice to disclose data is the duty of the distant agent, the interplay maintains its opaque nature.

Taken collectively, these tenets paint an image of a protocol attempting to be sensible, safe, versatile, and respectful of the unbiased nature of brokers. However rules on paper are one factor; how does A2A truly implement these concepts? To see that, we have to shift from the design philosophy to the nuts and bolts—the particular mechanisms and elements that make agent-to-agent communication work.

Key Mechanisms and Parts of A2A

Translating these rules into follow requires particular mechanisms. Central to enabling brokers to grasp one another throughout the A2A framework is the Agent Card. This part features as a standardized digital enterprise card for an AI agent, usually supplied as a metadata file. Its main function is to publicly declare what an agent is, what it might probably do, the place it may be reached, and how you can work together with it.

Right here’s a simplified instance of what an Agent Card would possibly appear to be, conveying the important data:

{
  "title": "StockInfoAgent",
  "description": "Supplies present inventory worth data.",
  "url": "http://stock-info.instance.com/a2a",
  "supplier": { "group": "ABCorp" },
  "model": "1.0.0",
  "expertise": [
    {
      "id": "get_stock_price_skill",
      "name": "Get Stock Price",
      "description": "Retrieves current stock price for a company"
    }
  ]
}

(shortened for brevity)

The Agent Card serves as the important thing connector between the totally different actors within the A2A protocol. A consumer—which could possibly be one other agent or maybe the appliance the consumer is interacting with—finds the Agent Card for the service it wants. It makes use of the small print from the cardboard, just like the URL, to contact the distant agent (server), which then performs the requested job with out exposing its inner strategies and sends again the outcomes in accordance with the A2A guidelines.

As soon as brokers are in a position to learn one another’s capabilities, A2A buildings their collaboration round finishing particular duties. A job represents the basic unit of labor requested by a consumer from a distant agent. Importantly, every job is stateful, permitting it to trace progress over time, which is crucial for dealing with operations which may not be instantaneous—aligning with A2A’s “async first” precept.

Communication associated to a job primarily makes use of messages. These carry the continuing dialogue, together with preliminary directions from the consumer, standing updates, requests for clarification, and even intermediate “ideas” from the agent. When the duty is full, the ultimate tangible outputs are delivered as artifacts, that are immutable outcomes like recordsdata or structured information. Each messages and artifacts are composed of a number of components, the granular items of content material, every with an outlined kind (like textual content or a picture).

This complete trade depends on normal net applied sciences like HTTP and customary information codecs, guaranteeing a broad basis for implementation and compatibility. By defining these core objects—job, message, artifact, and half—A2A gives a structured method for brokers to handle requests, trade data, and ship outcomes, whether or not the work takes seconds or hours.

Safety is, after all, a essential concern for any protocol aiming for enterprise adoption, and A2A addresses this instantly. Relatively than inventing completely new safety mechanisms, it leans closely on established practices. A2A aligns with requirements just like the OpenAPI specification for outlining authentication strategies and usually encourages treating brokers like different safe enterprise functions. This enables the protocol to combine into current company safety frameworks, equivalent to established id and entry administration (IAM) techniques for authenticating brokers, making use of current community safety guidelines and firewall insurance policies to A2A endpoints, or probably feeding A2A interplay logs into centralized safety data and occasion administration (SIEM) platforms for monitoring and auditing.

A core precept is preserving delicate credentials, equivalent to API keys or entry tokens, separate from the principle A2A message content material. Purchasers are anticipated to acquire these credentials via an unbiased course of. As soon as obtained, they’re transmitted securely utilizing normal HTTP headers, a standard follow in net APIs. Distant brokers, in flip, clearly state their authentication necessities—typically inside their Agent Playing cards—and use normal HTTP response codes to handle entry makes an attempt, signaling success or failure in a predictable method. This reliance on acquainted net safety patterns lowers the barrier to implementing safe agent interactions.

A2A additionally facilitates the creation of a distributed “interplay reminiscence” throughout a multi-agent system by offering a standardized protocol for brokers to trade and reference task-specific data, together with distinctive identifiers (taskId, sessionId), standing updates, message histories, and artifacts. Whereas A2A itself doesn’t retailer this reminiscence, it permits every taking part A2A consumer and server agent to take care of its portion of the general job context. Collectively, these particular person agent recollections, linked and synchronized via A2A’s structured communication, type the excellent interplay reminiscence of the complete multi-agent system, permitting for coherent and stateful collaboration on complicated duties.

So, in a nutshell, A2A is an try to carry guidelines and standardization to the quickly evolving world of brokers by defining how unbiased techniques can uncover one another, collaborate on duties (even long-running ones), and deal with safety utilizing well-trodden net paths, all whereas preserving their internal workings non-public. It’s centered squarely on agent-to-agent communication, attempting to resolve the issue of remoted digital employees unable to coordinate.

However getting brokers to speak to one another is just one piece of the interoperability puzzle going through AI builders right now. There’s one other normal gaining vital traction that tackles a associated but distinct problem: How do these refined AI functions work together with the surface world—the databases, APIs, recordsdata, and specialised features also known as “instruments”? This brings us to Anthropic’s Mannequin Context Protocol, or MCP.

MCP: Mannequin Context Protocol Overview

It wasn’t so way back, actually, that giant language fashions (LLMs), whereas spectacular textual content mills, had been typically mocked for his or her typically hilarious blind spots. Requested to do easy arithmetic, rely the letters in a phrase precisely, or inform you the present climate, and the outcomes could possibly be confidently delivered but utterly incorrect. This wasn’t only a quirk; it highlighted a elementary limitation: The fashions operated purely on the patterns realized from their static coaching information, disconnected from reside data sources or the power to execute dependable procedures. However nowadays are largely over (or so it appears)—state-of-the-art AI fashions are vastly simpler than their predecessors from only a yr or two in the past.

A key cause for the effectiveness of AI techniques (brokers or not) is their capacity to attach past their coaching information: interacting with databases and APIs, accessing native recordsdata, and using specialised exterior instruments. Equally to interagent communication, nonetheless, there are some arduous challenges that have to be tackled first.

Integrating these AI techniques with exterior “instruments” includes collaboration between AI builders, agent architects, device suppliers, and others. A major hurdle is that device integration strategies are sometimes tied to particular LLM suppliers (like OpenAI, Anthropic, or Google), and these suppliers deal with device utilization in a different way. Defining a device for one system requires a particular format; utilizing that very same device with one other system typically calls for a special construction.

Contemplate the next examples.

OpenAI’s API expects a operate definition structured this manner:

{
  "kind": "operate",
  "operate": {
    "title": "get_weather",
    "description": "Retrieves climate information ...",
    "parameters": {...}
  }
}

Whereas Anthropic’s API makes use of a special structure:

{
  "title": "get_weather",
  "description": "Retrieves climate information ...",
  "input_schema": {...}
}

This incompatibility means device suppliers should develop and keep separate integrations for every AI mannequin supplier they need to assist. If an agent constructed with Anthropic fashions wants sure instruments, these instruments should observe Anthropic’s format. If one other developer needs to make use of the identical instruments with a special mannequin supplier, they basically duplicate the combination effort, adapting definitions and logic for the brand new supplier.

Format variations aren’t the one problem; language obstacles additionally create integration difficulties. For instance, getting a Python-based agent to instantly use a device constructed round a Java library requires appreciable growth effort.

This integration problem is exactly what the Mannequin Context Protocol was designed to resolve. It gives a normal method for various AI functions and exterior instruments to work together.

Much like A2A, MCP operates utilizing two key components, beginning with the MCP server. This part is answerable for exposing the device’s performance. It comprises the underlying logic—perhaps Python code hitting a climate API or routines for information entry—developed in an appropriate language. Servers generally bundle associated capabilities, like file operations or database entry instruments. The second part is the MCP consumer. This piece sits contained in the AI software (the chatbot, agent, or coding assistant). It finds and connects to MCP servers which can be accessible. When the AI app or mannequin wants one thing from the surface world, the consumer talks to the appropriate server utilizing the MCP normal.

The hot button is that communication between consumer and server adheres to the MCP normal. This adherence ensures that any MCP-compatible consumer can work together with any MCP server, regardless of the consumer’s underlying AI mannequin or the language used to construct the server.

Adopting this normal gives a number of benefits:

• Construct as soon as, use wherever: Create a functionality as an MCP server as soon as; any MCP-supporting software can use it.
• Language flexibility: Develop servers within the language greatest fitted to the duty.
• Leverage ecosystem: Use current open supply MCP servers as a substitute of constructing each integration from scratch.
• Improve AI capabilities: Simply give brokers, chatbots, and assistants entry to various real-world instruments.

Adoption of MCP is accelerating, demonstrated by suppliers equivalent to GitHub and Slack, which now supply servers implementing the protocol.

MCP and A2A

However how do the Mannequin Context Protocol and the Agent2Agent (A2A) Protocol relate? Do they resolve the identical downside or serve totally different features? The strains can blur, particularly since many agent frameworks permit treating one agent as a device for an additional (agent as a device).

Each protocols enhance interoperability inside AI techniques, however they function at totally different ranges. By analyzing their variations in implementation and targets we will clearly establish key differentiators.

MCP focuses on standardizing the hyperlink between an AI software (or agent) and particular, well-defined exterior instruments or capabilities. MCP makes use of exact, structured schemas (like JSON Schema) to outline instruments, establishing a transparent API-like contract for predictable and environment friendly execution. For instance, an agent needing the climate would use MCP to name a get_weather device on an MCP climate server, specifying the placement “London.” The required enter and output are strictly outlined by the server’s MCP schema. This method removes ambiguity and solves the issue of incompatible device definitions throughout LLM suppliers for that particular operate name. MCP often includes synchronous calls, supporting dependable and repeatable execution of features (until, after all, the climate in London has modified within the meantime, which is completely believable).

A2A, alternatively, standardizes how autonomous brokers talk and collaborate. It excels at managing complicated, multistep duties involving coordination, dialogue, and delegation. Relatively than relying on inflexible operate schemas, A2A interactions make the most of pure language, making the protocol higher fitted to ambiguous targets or duties requiring interpretation. A very good instance could be “Summarize market traits for sustainable packaging.” Asynchronous communication is a key tenet of A2A, which additionally contains mechanisms to supervise the lifecycle of probably prolonged duties. This includes monitoring standing (like working, accomplished, and enter required) and managing the required dialogue between brokers. Contemplate a trip planner agent utilizing A2A to delegate book_flights and reserve_hotel duties to specialised journey brokers whereas monitoring their standing. In essence, A2A’s focus is the orchestration of workflows and collaboration between brokers.

This distinction highlights why MCP and A2A operate as complementary applied sciences, not rivals. To borrow an analogy: MCP is like standardizing the wrench a mechanic makes use of—defining exactly how the device engages with the bolt. A2A is like establishing a protocol for a way that mechanic communicates with a specialist mechanic throughout the workshop (“Listening to a rattle from the entrance left, are you able to diagnose?”), initiating a dialogue and collaborative course of.

In refined AI techniques, we will simply think about them working collectively: A2A would possibly orchestrate the general workflow, managing delegation and communication between totally different brokers, whereas these particular person brokers would possibly use MCP below the hood to work together with particular databases, APIs, or different discrete instruments wanted to finish their a part of the bigger job.

Placing It All Collectively

We’ve mentioned A2A for agent collaboration and MCP for device interplay as separate ideas. However their actual potential would possibly lie in how they work collectively. Let’s stroll via a easy, sensible state of affairs to see how these two protocols might operate in live performance inside a multi-agent system.

Think about a consumer asks their main interface agent—let’s name it the Host Agent—an easy query: “What’s Google’s inventory worth proper now?”

The Host Agent, designed for consumer interplay and orchestrating duties, doesn’t essentially know how you can fetch inventory costs itself. Nevertheless, it is aware of (maybe by consulting an agent registry through an Agent Card) a few specialised Inventory Information Agent that handles monetary information. Utilizing A2A, the Host Agent delegates the duty: It sends an A2A message to the Inventory Information Agent, basically saying, “Request: Present inventory worth for GOOGL.”

The Inventory Information Agent receives this A2A job. Now, this agent is aware of the particular process to get the information. It doesn’t want to debate it additional with the Host Agent; its job is to retrieve the worth. To do that, it turns to its personal toolset, particularly an MCP inventory worth server. Utilizing MCP, the Inventory Information Agent makes a exact, structured name to the server—successfully get_stock_price(image: "GOOGL"). This isn’t a collaborative dialogue just like the A2A trade; it’s a direct operate name utilizing the standardized MCP format.

The MCP server does its job: appears up the worth and returns a structured response, perhaps {"worth": "174.92 USD"}, again to the Inventory Information Agent through MCP.

With the information in hand, the Inventory Information Agent completes its A2A job. It sends a ultimate A2A message again to the Host Agent, reporting the outcome: "Outcome: Google inventory is 174.92 USD."

Lastly, the Host Agent takes this data obtained through A2A and presents it to the consumer.

Even on this easy instance, the complementary roles change into clear. A2A handles the higher-level coordination and delegation between autonomous brokers (Host delegates to Inventory Information). MCP handles the standardized, lower-level interplay between an agent and a particular device (Inventory Information makes use of the worth server). This creates a separation of considerations: The Host agent doesn’t must learn about MCP or inventory APIs, and the Inventory Information agent doesn’t must deal with complicated consumer interplay—it simply fulfills A2A duties, utilizing MCP instruments the place essential. Each brokers stay largely opaque to one another, interacting solely via the outlined protocols. This modularity, enabled through the use of each A2A for collaboration and MCP for device use, is vital to constructing extra complicated, succesful, and maintainable AI techniques.

Conclusion and Future Work

We’ve outlined the challenges of constructing AI brokers collaborate, explored Google’s A2A protocol as a possible normal for interagent communication, and in contrast and contrasted it with Anthropic’s Mannequin Context Protocol. Standardizing device use and agent interoperability are essential steps ahead in enabling efficient and environment friendly multi-agent system (MAS) design.

However the story is much from over, and agent discoverability is among the rapid subsequent challenges that have to be tackled. When speaking to enterprises it turns into manifestly apparent that that is typically very excessive on their precedence listing. As a result of, whereas A2A defines how brokers talk as soon as linked, the query of how they discover one another within the first place stays a big space for growth. Easy approaches may be carried out—like publishing an Agent Card at a normal net handle and capturing that handle in a listing—however that feels inadequate for constructing a really dynamic and scalable ecosystem. That is the place we see the idea of curated agent registries come into focus, and it’s maybe some of the thrilling areas of future work for MAS.

We think about an inner “agent retailer” (akin to an app retailer) or skilled itemizing for a corporation’s AI brokers. Builders might register their brokers, full with versioned expertise and capabilities detailed of their Agent Playing cards. Purchasers needing a particular operate might then question this registry, looking not simply by title however by required expertise, belief ranges, or different very important attributes. Such a registry wouldn’t simply simplify discovery; it could foster specialization, allow higher governance, and make the entire system extra clear and manageable. It strikes us from merely discovering an agent to discovering the proper agent for the job based mostly on its declared expertise.

Nevertheless, even refined registries can solely assist us discover brokers based mostly on these declared capabilities. One other fascinating, and maybe extra elementary, problem for the long run: coping with emergent capabilities. One of many outstanding facets of contemporary brokers is their capacity to mix various instruments in novel methods to deal with unexpected issues. An agent geared up with varied mapping, site visitors, and occasion information instruments, as an illustration, might need “route planning” listed on its Agent Card. However by creatively combining these instruments, it may also be able to producing complicated catastrophe evacuation routes or extremely customized multistop itineraries—essential capabilities seemingly unlisted just because they weren’t explicitly predefined. How can we reconcile the necessity for predictable, discoverable expertise with the highly effective, adaptive problem-solving that makes brokers so promising? Discovering methods for brokers to sign or for shoppers to find these unlisted potentialities with out sacrificing construction is a big open query for the A2A group and the broader area (as highlighted in discussions like this one).

Addressing this problem provides one other layer of complexity when envisioning future MAS architectures. Trying down the street, particularly inside giant organizations, we’d see the registry concept evolve into one thing akin to the “information mesh” idea—a number of, probably federated registries serving particular domains. This might result in an “agent mesh”: a resilient, adaptable panorama the place brokers collaborate successfully below a unified centralized governance layer and distributed administration capabilities (e.g., introducing notions of a knowledge/agent steward who manages the standard, accuracy, and compliance of a enterprise unit information/brokers). However guaranteeing this mesh can leverage each declared and emergent capabilities shall be key. Exploring that totally, nonetheless, is probably going a subject for an additional day.

Finally, protocols like A2A and MCP are very important constructing blocks, however they’re not the complete map. To construct multi-agent techniques which can be genuinely collaborative and sturdy, we’d like extra than simply normal communication guidelines. It means stepping again and considering arduous in regards to the general structure, wrestling with sensible complications like safety and discovery (each the specific sort and the implicit, emergent kind), and acknowledging that these requirements themselves must adapt as we be taught. The journey from right now’s often-siloed brokers to actually cooperative ecosystems is ongoing, however initiatives like A2A supply useful markers alongside the way in which. It’s undoubtedly a tricky engineering street forward. But, the prospect of AI techniques that may actually work collectively and deal with complicated issues in versatile methods? That’s a vacation spot well worth the effort.