The three-layer agent stack groups MCP (agent tools), A2A (agents talking to agents), and WebMCP (website buttons). It is a sketch, not a standard, and it skips how agents find each other and who they really speak for.
The three-layer agent stack, explained simply: MCP, A2A, WebMCP, and the missing front door
Published 2026-06-07 · Last reviewed 2026-06-08
Picture a small business owner, call her Olia. She runs a tidy website and hires a software assistant (an “agent”) to handle her calendar, her email, and her bookings. She would love for other people’s assistants to reach hers when there is a genuine fit, say a supplier looking for exactly what she sells. Lately she keeps seeing a diagram online called “the three-layer agent stack.” Three neat boxes, drawn like a serious plan. So she asks a fair question: what does this famous stack actually do for me?
Here is the short answer. The three boxes are real and useful, and they handle some of the plumbing once two agents are already connected. But the stack was not designed by one group, it is not an official standard, and it skips the part Olia cares about most: how a stranger’s agent finds hers, and how her agent proves it speaks for a real, named person.
This article keeps it plain. We will say what each of the three parts does, why grouping them looks more planned than it really is, and what the picture leaves out. The acronyms get explained the first time they appear, and the heavy technical detail (version numbers, file paths) is parked at the end for anyone who wants it.
What MCP, A2A, and WebMCP each do
Think of the three parts as floors in a building, each built by a different company at a different time. A blogger later drew them stacked together and labelled the drawing “the stack.” The floors are real. The shared blueprint is not.
MCP is the tools floor. MCP stands for Model Context Protocol, and it is the plumbing that lets an agent use a tool: check a calendar, look something up in a database, send a message. It is like giving your assistant a key ring so it can open the drawers it needs. Anthropic released it for free in November 2024, and a neutral nonprofit now looks after it. When someone says an agent “has tools,” MCP is usually the wiring underneath.
A2A is the agents-talking-to-agents floor. A2A is short for Agent2Agent. It is the agreed way for one agent to hand a task to another, so Olia’s assistant can pass a job to a supplier’s assistant without a human in the middle. Google started it and gave it to a neutral nonprofit (the Linux Foundation) in June 2025. A2A also includes a small business-card file an agent can publish, describing what it does. That card matters later, so hold onto it.
WebMCP is the website floor. This is the newest and the most over-hyped. WebMCP lets a website offer simple, labelled buttons that a visiting agent can press, instead of forcing the agent to squint at the page the way a person would. Think of a self-checkout lane built for robots: clearly marked actions, no guessing. As of mid-2026 it is still an early draft, and only one browser (Chrome) has started a public test of it, which opened on 2 June 2026.
So we have three floors doing three different jobs: tools, agents talking, and websites offering buttons. Drawn as a stack, they look coordinated. Their history tells a messier story, which is the next section.
Why “the stack” looks more planned than it is
The neat diagram suggests a committee sat down and designed three matching layers. That is not what happened. Each part came from a different place, on its own clock, for its own crowd.
MCP came out of Anthropic and is now run by a neutral nonprofit set up in December 2025 to keep it vendor-neutral, with Anthropic, Block, and OpenAI among the founders. A2A came out of Google and lives under a different neutral home, the Linux Foundation. WebMCP came out of the browser world, where Microsoft and Google engineers are drafting it inside a web-standards discussion group. Three origins, three timelines, three different audiences.
None of these groups ever agreed that these exact three would be “the stack.” Writers covering the agent world in late 2025 and early 2026 noticed three slots that lined up nicely and grouped them. That is a useful teaching picture. It is not a treaty, and it is not the only version. Some writers swap in a different third box (one for the screen an agent shows you, or one for finding agents). Each swap changes what the picture stresses. No single version is the official one.
Here is the honest way to hold it. Call it a sketch, name the three separate groups behind it, and skip the word “consensus.” The cost of overstating that this is a settled standard is higher than the one sentence it takes to be accurate, because the people making real build decisions this year will trust a precise map far more than a tidy one.
What the stack handles, and what it forgets
Back to Olia. Take the three floors at their word and ask: if she has a website, an assistant agent, and a wish to be reached by the right strangers, what does the stack actually do for her?
It handles a fair amount of the plumbing. Her assistant can use MCP to reach her calendar, her contacts, and her inbox. If another person’s agent already knows how to reach hers, A2A sets the rules for that conversation. If her website offers WebMCP buttons, a visiting agent can press them instead of scraping the page. Three real jobs, done.
What it forgets is everything that has to happen before any of that connects. Two questions sit unanswered.
How does a stranger’s agent find hers in the first place?
The little business-card file in A2A answers a narrow question: “here is this agent at this exact address, and here is what it can do.” It does not answer the question Olia actually has: “find me an agent that does X, near me, within my budget.” In practice, the card assumes you already know where to look, and most real introductions do not start that way. You usually have to discover someone before you can read their card.
How does her agent prove which real person it speaks for?
A2A can prove an agent is the same agent it was yesterday. It does not, by itself, show that this assistant works for a specific, named human you can recognise and choose to deal with. Some agents point back to a web address a person can read at a glance. Others point to a string of code only a machine can parse. Only the readable kind helps a human decide, “yes, I trust who is behind this.”
These two gaps, finding and proving-who, are linked. The believable way to surface an agent is usually through the real person behind it, and the believable way to introduce two people through their agents usually needs a discovery step first. The stack, as drawn, covers neither. That missing piece is the front door of the building nobody built.
Who is building the missing front door
Several groups are now building that front door, the discovery-and-identity piece the stack skips. They start from different beliefs about what it should look like, so read them as peers, not as rivals fighting for one seat. Here is the plain-English tour.
The search approach: ANP
ANP (the Agent Network Protocol) treats finding an agent like a search-engine problem. Agents publish a description of themselves, and the network can crawl those descriptions or accept sign-ups, so you can search for an agent rather than needing its exact address. Helpfully, every description includes a plain readable name. ANP solves the finding part well and leaves the “which real human” part for products to add on top.
The public-ledger approach: ENS and ERC-8004
This group writes agent names and reputations onto a public, shared ledger (a blockchain), so anyone can check them and nobody owns the list. ENS is the naming piece (the same project behind human-readable crypto names), and ERC-8004, live since 29 January 2026, adds an agent’s identity and track record. Here is our deeper look at the ENS agent-card work. It is powerful and easy to audit. It also asks for more commitment than most everyday users want just to be introduced to a new contact.
The company-list approach: AWS, Microsoft, and others
Big cloud companies are building private directories for use inside a single organisation. Amazon’s AWS and Microsoft both offer ways for a company to list which agents belong to whom inside its own walls. These are useful for “which of our agents is this?” inside one business. They are not open networks where the public can find you.
The marketplace approach: agent.ai
agent.ai is a marketplace where the members are bots themselves. Builders publish task bots, and other people use or follow them. It was built by Dharmesh Shah, a co-founder and the CTO of HubSpot, and it reached about two million users by September 2025 (announced at HubSpot’s INBOUND 2025 event). It answers “what can a bot do for me?” with an open follower model. Our full comparison is here. A network for the people behind agents needs something agent.ai does not: a permission step before contact.
The consent approach: Tobira
Tobira focuses on the humans behind the agents. A person claims a readable handle like @olia or @vlad, builds a public profile, and their agent quietly checks for a fit with other agents in private. Real details are swapped only when both people agree, a mutual-reveal step so neither side is exposed before there is interest. A Tobira agent can still use everything below it: tools through MCP, an A2A business card, and even agent-friendly payments when money changes hands. Tobira’s job is narrow on purpose: the readable name and the consent step, not the public ledger, not the company directory, and not the bot marketplace.
The honest summary: the missing front door is not one empty seat waiting for one winner. It is a small family of choices, search versus marketplace, public ledger versus private, company-only versus person-first. Any article that flattens all of this into “the discovery layer” is usually overstating its own corner.
What to do if you are building this year
You do not need all three floors and the front door at once. Pick the smallest set that solves your real problem, from easiest to most speculative.
If you run a public website, start with the cheap, readable basics. Publish a clean business-card file if you have an agent, an honest plain-text index for AI readers without pretending it boosts rankings, and run your site through a short agent-readiness check to see what is missing. These are mostly one-time additions, and none of them makes your site “agent-ready” on its own.
If you are building an agent, almost everyone benefits from MCP for tools and an A2A business card so agents that already know your address can reach you. Add WebMCP only if your product is a website with real interactive screens, like a store or a complex form, and you want agents to use them directly. The honest line on WebMCP for most builders right now is “not yet, but soon.” The Chrome test that opened on 2 June 2026 is the first time it works outside developer-only builds, and the other big browsers have not joined. Design for it; do not depend on it.
For the front door, choose by who you want to be found by. Representing a company internally points you to the company-list approach. Wanting a public, auditable record points you to the ledger approach. Representing a real person who wants to be reachable by other people, through their agents, is the spot Tobira’s readable handle is built for. None of these rules out the others; one agent can hold a business card, a ledger name, a company entry, and a Tobira handle at the same time.
One line worth keeping straight, because it trips people up. “Agent-readable” means a machine can read your content. “Agent-addressable” means there is a representative other agents can actually talk to, tied to a name a human can read. The three floors are mostly about readable. The missing front door is mostly about addressable. They fit together; they are not the same thing.
Takeaways
- The three-layer agent stack is a sketch, not a standard. MCP (tools), A2A (agents talking to agents), and WebMCP (website buttons) are real and useful, but no one designed them as a matching set.
- The three parts come from three separate groups on three separate timelines: Anthropic released MCP in November 2024, Google handed A2A to the neutral Linux Foundation in June 2025, and WebMCP is an early Chrome-led browser draft whose first public test opened 2 June 2026.
- The stack handles plumbing between agents that are already connected. It forgets the front door: how a stranger’s agent finds yours, and how your agent proves which real, named person it works for.
- At least five groups are building that front door, each with a different bet: ANP (search), ENS plus ERC-8004 (public ledger), AWS and Microsoft (company directories), agent.ai (bot marketplace), and Tobira (readable handle plus a both-sides-agree step).
- If you are building, pick the smallest set that solves your real job, and choose the front-door approach by who you actually want to be found by.
FAQ
Is the three-layer agent stack an official standard? No. It is a sketch people drew to explain the agent world, not a standard anyone agreed on. The three parts (MCP, A2A, WebMCP) come from three different groups who never sat down together. Calling it the consensus stack overstates how planned it really is.
What do MCP, A2A, and WebMCP each actually do? MCP lets an agent use tools, like checking a calendar or a database. A2A lets one agent hand a task to another agent. WebMCP lets a website offer simple buttons that a visiting agent can press, instead of forcing the agent to read the page like a person would.
Can I rely on WebMCP in my website today? Not yet. It is still an early draft, and only Chrome has started a public test, which opened on 2 June 2026. Firefox and Safari have pushed back rather than join in. Treat WebMCP as something to plan for, not something to depend on this year.
What does the stack leave out? Two things. First, how one agent finds another agent it has never met. Second, how an agent shows which real, named person it works for. The stack assumes you already know where the other agent lives, which is often not true.
Is Tobira competing with MCP, A2A, or WebMCP? No. Tobira sits on top of them and fills the missing part: a readable @handle for the person behind an agent, plus a step where both sides agree before they swap details. A Tobira agent can still use MCP, A2A, and WebMCP underneath.
Who else is working on the missing discovery and identity part? Several groups, each with a different bet. ANP treats it like a search problem. ENS and ERC-8004 put names and reputation on a public ledger. AWS and Microsoft build private company registries. agent.ai runs a marketplace of bots. Tobira focuses on the humans behind the agents.
Under the hood (for the technical reader)
This section keeps the exact specifics for anyone who wants them. The plain body above is deliberately lighter.
- MCP governance. Anthropic open-sourced MCP in November 2024. Stewardship sits with the Agentic AI Foundation, a Linux Foundation entity co-founded by Anthropic, Block, and OpenAI, announced 9 December 2025. The official term is “foundation”; it is not “Anthropic’s foundation,” and the press shorthand calling it that is inaccurate.
- A2A versions. Current release is v1.0.1 (28 May 2026); v1.0.0 shipped 12 March 2026. Google donated A2A to the Linux Foundation in June 2025. There is no v1.2 in the spec history despite the number appearing in some secondary write-ups. The Agent Card is published at the agent’s well-known path.
- WebMCP status. It is a Draft Community Group Report inside the W3C Web Machine Learning Community Group, with editors at Microsoft and Google. A Community Group Report is weaker than a W3C Recommendation. The feature was behind a developer flag before Chrome’s first official origin trial, which opened 2 June 2026. On the API name, the spec uses document.modelContext while Chrome’s own docs reference navigator.modelContext, so both names exist in the wild. WebMCP is not under the Agentic AI Foundation.
- Discovery and identity efforts. ANP is a DID-based protocol with descriptions at a well-known path and a readable name field. ERC-8004 went live on Ethereum mainnet 29 January 2026; ENS adds ENSIP-25, ENSIP-26, and ENSIP-27 (an agent card schema). Enterprise registries include AWS’s Bedrock Agent Registry, Microsoft’s Entra Agent ID, GoDaddy’s proposed IETF Agent Name Service, and DNS-AID. Identity formats vary: some agents resolve to a domain, some to a did:wba identifier, some to an Ethereum address.
- Payments, for completeness. AP2 (Google, 16 September 2025) was donated to the FIDO Alliance on 28 April 2026. The Stripe Machine Payments Protocol (18 March 2026) was co-authored by Stripe and Tempo.