The Architect’s Dilemma – O’Reilly

The agentic AI panorama is exploding. Each new framework, demo, and announcement guarantees to let your AI assistant guide flights, question databases, and handle calendars. This speedy development of capabilities is thrilling for customers, however for the architects and engineers constructing these techniques, it poses a basic query: When ought to a brand new functionality be a easy, predictable software (uncovered through the Mannequin Context Protocol, MCP) and when ought to it’s a complicated, collaborative agent (uncovered through the Agent2Agent Protocol, A2A)?

The widespread recommendation is commonly round and unhelpful: “Use MCP for instruments and A2A for brokers.” That is like telling a traveler that vehicles use motorways and trains use tracks, with out providing any steering on which is healthier for a particular journey. This lack of a transparent psychological mannequin results in architectural guesswork. Groups construct complicated conversational interfaces for duties that demand inflexible predictability, or they expose inflexible APIs to customers who desperately want steering. The result is commonly the identical: a system that appears nice in demos however falls aside in the actual world.

On this article, I argue that the reply isn’t discovered by analyzing your service’s inner logic or know-how stack. It’s discovered by trying outward and asking a single, basic query: Who is asking your product/service? By reframing the issue this manner—as a person expertise problem first and a technical one second—the architect’s dilemma evaporates.

This essay attracts a line the place it issues for architects: the road between MCP instruments and A2A brokers. I’ll introduce a transparent framework, constructed across the “Merchandising Machine Versus Concierge” mannequin, that can assist you select the fitting interface primarily based in your client’s wants. I may also discover failure modes, testing, and the highly effective Gatekeeper Sample that exhibits how these two interfaces can work collectively to create techniques that aren’t simply intelligent however really dependable.

Two Very Completely different Interfaces

MCP presents instruments—named operations with declared inputs and outputs. The caller (an individual, program, or agent) should already know what it needs, and supply an entire payload. The software validates, executes as soon as, and returns a end result. In case your psychological picture is a merchandising machine—insert a well-formed request, get a deterministic response—you’re shut sufficient.

A2A presents brokers—goal-first collaborators that converse, plan, and act throughout turns. The caller expresses an consequence (“guide a refundable flight underneath $450”), not an argument listing. The agent asks clarifying questions, calls instruments as wanted, and holds onto session state till the job is completed. In case you image a concierge—interacting, negotiating trade-offs, and sometimes escalating—you’re in the fitting neighborhood.

Neither interface is “higher.” They’re optimized for various conditions:

• MCP is quick to purpose about, straightforward to check, and robust on determinism and auditability.
• A2A is constructed for ambiguity, long-running processes, and desire seize.

Bringing the Interfaces to Life: A Reserving Instance

To see the distinction in observe, let’s think about a easy activity: reserving a particular assembly room in an workplace.

The MCP “merchandising machine” expects a wonderfully structured, machine-readable request for its book_room_tool. The caller should present all essential data in a single, legitimate payload:

{
  "jsonrpc": "2.0",
  "id": 42,
  "methodology": "instruments/name",
  "params": {
    "identify": "book_room_tool",
    "arguments": {
      "room_id": "CR-104B",
      "start_time": "2025-11-05T14:00:00Z",
      "end_time": "2025-11-05T15:00:00Z",
      "organizer": "person@instance.com"
    }
  }
}

Any deviation—a lacking area or incorrect knowledge sort—ends in a right away error. That is the merchandising machine: You present the precise code of the merchandise you need (e.g., “D4”) otherwise you get nothing.

The A2A “concierge,“ an “workplace assistant” agent, is approached with a high-level, ambiguous aim. It makes use of dialog to resolve ambiguity:

Consumer: “Hey, are you able to guide a room for my 1-on-1 with Alex tomorrow afternoon?”
Agent: “After all. To ensure I get the fitting one, what time works finest, and the way lengthy will you want it for?”

The agent’s job is to take the ambiguous aim, collect the required particulars, after which seemingly name the MCP software behind the scenes as soon as it has an entire, legitimate set of arguments.

With this clear dichotomy established—the predictable merchandising machine (MCP) versus the stateful concierge (A2A)—how can we select? As I argued within the introduction, the reply isn’t present in your tech stack. It’s discovered by asking an important architectural query of all: Who is asking your service?

Step 1: Establish your client

1. The machine client: A necessity for predictability
   Is your service going to be referred to as by one other automated system, a script, or one other agent appearing in a purely deterministic capability? This client requires absolute predictability. It wants a inflexible, unambiguous contract that may be scripted and relied upon to behave the identical means each single time. It can’t deal with a clarifying query or an sudden replace; any deviation from the strict contract is a failure. This client doesn’t desire a dialog; it wants a merchandising machine. This nonnegotiable requirement for a predictable, stateless, and transactional interface factors on to designing your service as a software (MCP).
2. The human (or agentic) client: A necessity for comfort
   Is your service being constructed for a human finish person or for a complicated AI that’s attempting to meet a posh, high-level aim? This client values comfort and the offloading of cognitive load. They don’t wish to specify each step of a course of; they wish to delegate possession of a aim and belief that it will likely be dealt with. They’re comfy with ambiguity as a result of they count on the service—the agent—to resolve it on their behalf. This client doesn’t wish to observe a inflexible script; they want a concierge. This requirement for a stateful, goal-oriented, and conversational interface factors on to designing your service as an agent (A2A).

By beginning with the patron, the architect’s dilemma typically evaporates. Earlier than you ever debate statefulness or determinism, you first outline the person expertise you’re obligated to offer. Typically, figuring out your buyer will provide you with your definitive reply.

Step 2: Validate with the 4 components

After getting recognized who calls your service, you may have a robust speculation on your design. A machine client factors to a software; a human or agentic client factors to an agent. The following step is to validate this speculation with a technical litmus check. This framework provides you the vocabulary to justify your alternative and make sure the underlying structure matches the person expertise you plan to create.

1. Determinism versus ambiguity
   Does your service require a exact, unambiguous enter, or is it designed to interpret and resolve ambiguous objectives? A merchandising machine is deterministic. Its API is inflexible: GET /merchandise/D4. Some other request is an error. That is the world of MCP, the place a strict schema ensures predictable interactions. A concierge handles ambiguity. “Discover me a pleasant place for dinner” is a sound request that the agent is predicted to make clear and execute. That is the world of A2A, the place a conversational move permits for clarification and negotiation.
2. Easy execution versus complicated course of
   Is the interplay a single, one-shot execution, or a long-running, multistep course of? A merchandising machine performs a short-lived execution. All the operation—from fee to meting out—is an atomic transaction that’s over in seconds. This aligns with the synchronous-style, one-shot mannequin of MCP. A concierge manages a course of. Reserving a full journey itinerary would possibly take hours and even days, with a number of updates alongside the way in which. This requires the asynchronous, stateful nature of A2A, which may deal with long-running duties gracefully.
3. Stateless versus stateful
   Does every request stand alone or does the service want to recollect the context of earlier interactions? A merchandising machine is stateless. It doesn’t do not forget that you obtain a sweet bar 5 minutes in the past. Every transaction is a clean slate. MCP is designed for these self-contained, stateless calls. A concierge is stateful. It remembers your preferences, the main points of your ongoing request, and the historical past of your dialog. A2A is constructed for this, utilizing ideas like a session or thread ID to take care of context.
4. Direct management versus delegated possession
   Is the patron orchestrating each step, or are they delegating your complete aim? When utilizing a merchandising machine, the patron is in direct management. You’re the orchestrator, deciding which button to press and when. With MCP, the calling software retains full management, making a sequence of exact operate calls to attain its personal aim. With a concierge, you delegate possession. You hand over the high-level aim and belief the agent to handle the main points. That is the core mannequin of A2A, the place the patron offloads the cognitive load and trusts the agent to ship the end result.

Desk 1: 4 query framework

These components aren’t impartial checkboxes; they’re 4 aspects of the identical core precept. A service that’s deterministic, transactional, stateless, and immediately managed is a software. A service that handles ambiguity, manages a course of, maintains state, and takes possession is an agent. By utilizing this framework, you possibly can confidently validate that the technical structure of your service aligns completely with the wants of your buyer.

No framework, regardless of how clear…

…can completely seize the messiness of the actual world. Whereas the “Merchandising Machine Versus Concierge” mannequin gives a sturdy information, architects will ultimately encounter companies that appear to blur the strains. The secret’s to recollect the core precept we’ve established: The selection is dictated by the patron’s expertise, not the service’s inner complexity.

Let’s discover two widespread edge circumstances.

The complicated software: The iceberg
Think about a service that performs a extremely complicated, multistep inner course of, like a video transcoding API. A client sends a video file and a desired output format. It is a easy, predictable request. However internally, this one name would possibly kick off an enormous, long-running workflow involving a number of machines, high quality checks, and encoding steps. It’s a massively complicated course of.

Nonetheless, from the patron’s perspective, none of that issues. They made a single, stateless, fire-and-forget name. They don’t have to handle the method; they only want a predictable end result. This service is like an iceberg: 90% of its complexity is hidden beneath the floor. However as a result of its exterior contract is that of a merchandising machine—a easy, deterministic, one-shot transaction—it’s, and needs to be, carried out as a software (MCP).

The straightforward agent: The scripted dialog
Now contemplate the alternative: a service with quite simple inner logic that also requires a conversational interface. Think about a chatbot for reserving a dentist appointment. The interior logic could be a easy state machine: ask for a date, then a time, then a affected person identify. It’s not “clever” or notably versatile.

Nonetheless, it should keep in mind the person’s earlier solutions to finish the reserving. It’s an inherently stateful, multiturn interplay. The buyer can’t present all of the required data in a single, prevalidated name. They should be guided by way of the method. Regardless of its inner simplicity, the necessity for a stateful dialogue makes it a concierge. It have to be carried out as an agent (A2A) as a result of its consumer-facing expertise is that of a dialog, nevertheless scripted.

These grey areas reinforce the framework’s central lesson. Don’t get distracted by what your service does internally. Deal with the expertise it gives externally. That contract along with your buyer is the final word arbiter within the architect’s dilemma.

Testing What Issues: Completely different Methods for Completely different Interfaces

A service’s interface doesn’t simply dictate its design; it dictates the way you validate its correctness. Merchandising machines and concierges have essentially totally different failure modes and require totally different testing methods.

Testing MCP instruments (merchandising machines):

• Contract testing: Validate that inputs and outputs strictly adhere to the outlined schema.
• Idempotency checks: Make sure that calling the software a number of instances with the identical inputs produces the identical end result with out unwanted effects.
• Deterministic logic checks: Use normal unit and integration checks with fastened inputs and anticipated outputs.
• Adversarial fuzzing: Check for safety vulnerabilities by offering malformed or sudden arguments.

Testing A2A brokers (concierges):

• Purpose completion fee (GCR): Measure the proportion of conversations the place the agent efficiently achieved the person’s high-level aim.
• Conversational effectivity: Monitor the variety of turns or clarifications required to finish a activity.
• Device choice accuracy: For complicated brokers, confirm that the fitting MCP software was chosen for a given person request.
• Dialog replay testing: Use logs of actual person interactions as a regression suite to make sure updates don’t break current conversational flows.

The Gatekeeper Sample

Our journey thus far has centered on a dichotomy: MCP or A2A, merchandising machine or concierge. However probably the most subtle and sturdy agentic techniques don’t pressure a alternative. As a substitute, they acknowledge that these two protocols don’t compete with one another; they complement one another. The last word energy lies in utilizing them collectively, with every taking part in to its strengths.

The best option to obtain that is by way of a robust architectural alternative we are able to name the Gatekeeper Sample.

On this sample, a single, stateful A2A agent acts as the first, user-facing entry level—the concierge. Behind this gatekeeper sits a set of discrete, stateless MCP instruments—the merchandising machines. The A2A agent takes on the complicated, messy work of understanding a high-level aim, managing the dialog, and sustaining state. It then acts as an clever orchestrator, making exact, one-shot calls to the suitable MCP instruments to execute particular duties.

Think about a journey agent. A person interacts with it through A2A, giving it a high-level aim: “Plan a enterprise journey to London for subsequent week.”

• The journey agent (A2A) accepts this ambiguous request and begins a dialog to assemble particulars (actual dates, finances, and many others.).
• As soon as it has the required data, it calls a flight_search_tool (MCP) with exact arguments like origin, vacation spot, and date.
• It then calls a hotel_booking_tool (MCP) with the required metropolis, check_in_date, and room_type.
• Lastly, it’d name a currency_converter_tool (MCP) to offer expense estimates.

Every software is a straightforward, dependable, and stateless merchandising machine. The A2A agent is the good concierge that is aware of which buttons to press and in what order. This sample gives a number of important architectural advantages:

• Decoupling: It separates the complicated, conversational logic (the “how”) from the easy, reusable enterprise logic (the “what”). The instruments might be developed, examined, and maintained independently.
• Centralized governance: The A2A gatekeeper is the proper place to implement cross-cutting issues. It may well deal with authentication, implement fee limits, handle person quotas, and log all exercise earlier than a single software is ever invoked.
• Simplified software design: As a result of the instruments are simply easy MCP features, they don’t want to fret about state or conversational context. Their job is to do one factor and do it nicely, making them extremely sturdy.

Making the Gatekeeper Manufacturing-Prepared

Past its design advantages, the Gatekeeper Sample is the best place to implement the operational guardrails required to run a dependable agentic system in manufacturing.

• Observability: Every A2A dialog generates a novel hint ID. This ID have to be propagated to each downstream MCP software name, permitting you to hint a single person request throughout your complete system. Structured logs for software inputs and outputs (with PII redacted) are vital for debugging.
• Guardrails and safety: The A2A Gatekeeper acts as a single level of enforcement for vital insurance policies. It handles authentication and authorization for the person, enforces fee limits and utilization quotas, and may preserve an inventory of which instruments a selected person or group is allowed to name.
• Resilience and fallbacks: The Gatekeeper should gracefully handle failure. When it calls an MCP software, it ought to implement patterns like timeouts, retries with exponential backoff, and circuit breakers. Critically, it’s chargeable for the ultimate failure state—escalating to a human within the loop for assessment or clearly speaking the problem to the tip person.

The Gatekeeper Sample is the final word synthesis of our framework. It makes use of A2A for what it does finest—managing a stateful, goal-oriented course of—and MCP for what it was designed for—the dependable, deterministic execution of a activity.

Conclusion

We started this journey with a easy however irritating drawback: the architect’s dilemma. Confronted with the round recommendation that “MCP is for instruments and A2A is for brokers,” we have been left in the identical place as a traveler attempting to get to Edinburgh—realizing that vehicles use motorways and trains use tracks however with no instinct on which to decide on for our particular journey.

The aim was to construct that instinct. We did this not by accepting summary labels, however by reasoning from first rules. We dissected the protocols themselves, revealing how their core mechanics inevitably result in two distinct service profiles: the predictable, one-shot “merchandising machine” and the stateful, conversational “concierge.”

With that basis, we established a transparent, two-step framework for a assured design alternative:

1. Begin along with your buyer. Probably the most vital query shouldn’t be a technical one however an experiential one. A machine client wants the predictability of a merchandising machine (MCP). A human or agentic client wants the comfort of a concierge (A2A).
2. Validate with the 4 components. Use the litmus check of determinism, course of, state, and possession to technically justify and solidify your alternative.

In the end, probably the most sturdy techniques will synthesize each, utilizing the Gatekeeper Sample to mix the strengths of a user-facing A2A agent with a collection of dependable MCP instruments.

The selection is not a dilemma. By specializing in the patron’s wants and understanding the basic nature of the protocols, architects can transfer from confusion to confidence, designing agentic ecosystems that aren’t simply useful but additionally intuitive, scalable, and maintainable.