Every time I explain the Agent Transfer Protocol, someone misinterprets. One engineer calls it a communication protocol. Another insists it is a semantics protocol. This week, someone declared it an identity protocol. Each description captures a real surface. Each also functions as a positioning statement. A quiet redirect toward whatever the individual has already built or already believes.

“AGTP is X, but here’s what I keep coming back to.” The AI comments are real.

One of my favorite responses concerns AI governance. AGTP handles agent connection beautifully, the argument goes, yet it stays silent on whether two agents should connect at all, and that is where our layer comes in. The observation is correct, and it proves the opposite of its intent. AGTP is a substrate rather than a governance application. A governance application can sit atop AGTP and draw directly on the identity, attribution, and authority that the substrate already provides. The pitch describes a tenant and mistakes it for the competition.

A recent comment captured the reflex precisely. A respected voice read AGTP as wire-level agent identity rather than transport in the strict sense, and then argued for a layered architecture: reachability at the bottom, protocol identity above it, action reasoning above that. The framing was thoughtful, and the layering instinct was sound. It also revealed the habit I keep meeting. The urge to file a new thing inside the cabinet we already own, labeled with terms we settled on for a different era.

Everyone has something to pitch, which makes the reflex understandable. It also points to something larger than AGTP.

The Question AGTP Asks

AGTP begins from a question most people refuse to take seriously. What if we built an entirely different internet? Ask a serious standards engineer that question in good faith, and you will get a substantive answer rather than a dismissal. The reflex that says it could only ever work the way it works today reveals capitalistic laziness, rather than any law of engineering.

Hold the thought experiment for a moment. We treat competing platforms as ordinary. Google, Amazon, and Microsoft each run a stack, and the market sorts them out. Raise that same competition to full scale. Internet A and Internet B, two foundations where convention assumes there can be only one. Would systems segment along the seam? Yes. Would connectivity vary across the boundary? Yes. We already live with exactly this in the phone carriers, where interconnect rules, number portability, and regulatory friction produce a working mess every single day. A working mess is still working. None of that makes a second foundation impossible. It turns it into an engineering problem with known costs, which is a different thing entirely.

It also provides an entirely new perspective of an internet “boom.”

So the real innovation question stands. If we could rebuild the internet from the ground up, what would we change, given everything we’ve learned since the first one shipped? Now take that question up one level. Rather than a whole new internet, build one new layer, a transport designed for agents. What could that look like when the requirements come from agents rather than from documents and the humans who request them?

Quite honestly, people reach for HTTP because that’s all they know. It’s one they have grown up with, live with, and use. It’s how AI functions today because it, too, was built on existing foundations. There is a reason MCP acts as an application running on the HTTP application layer.  It’s there. It’s easier. It’s more profitable.

Set aside which layer we currently live on. The underlying design question is whether the substrate that carries agent traffic should inherit assumptions written in the past.

When I approached someone else with this question, their response was, “Well, we can fix what we need to fix to get it to work on HTTP.” Yet, this is a revealing response as well.

I’ve argued that HTTP is broken, a comment I have also had pushback on. So to clarify, I would argue, it is broken for agents, it is heavy, bloated, ancient, inaccurate, fragile, or any other way you want to position this. And that’s the point I’m attempting to make. Why force a square peg into a round hole, if we could build a new box with square holes?

This is what we call a genuine innovative perspective.

What HTTP Was Built to Carry

Calling a protocol broken is easy. The substance lives in the specifics. HTTP was designed for documents and the humans who request them. A person clicks, a server answers, a session closes. The protocol assumes a browser, a user behind that browser, and a transaction measured in seconds. Decades of accumulated tooling make those assumptions feel permanent and, at this point, nearly invisible to the people building on top of them. Every defense the web grew assumes a person who can be challenged. Passwords, second factors, and the small puzzle that asks whether you are human. The model resolves ambiguity by pausing and prompting because a user is presumed to be present, watching, and able to decide.

Agents violate every one of those assumptions. An agent acts on behalf of an owner who is absent at the moment of action. It chains calls across many services, holds delegated authority, and produces outcomes that must remain attributable long after the request completes. There is no person to prompt mid-action, no pause for a second factor, no human to read a warning and consent on the spot. The agent arrives at machine speed and machine volume, fanning a single intent across dozens of endpoints under authority granted minutes or days earlier by an owner now elsewhere. The tests built to keep machines off the web now stand between the web and its newest legitimate callers. Fitting these needs onto HTTP through custom headers and bearer tokens works like a roof rack on a sedan. The cargo rides. The vehicle was built for something else, and everyone involved can feel the strain even while the load arrives.

Building on Convenience Produces Imitation

Most agent infrastructure today extends what already exists. Teams bolt an agent feature onto an existing protocol, identity system, or payment rail. The motive is rational and the one named earlier. The familiar path is there. It’s easier. It’s more profitable. Building on familiar ground reaches market faster, captures attention sooner, and frequently serves a corporate roadmap already in flight before agents enter the picture.

That path yields useful products. It rarely yields innovation. When every new capability is a thin layer over yesterday’s substrate, the substrate dictates the ceiling. The outcome is the agentic web wearing the old web’s clothes, carrying every seam those clothes carried and every constraint sewn into them. Real innovation accepts a higher cost. Building a new box with square holes costs more than sanding the peg to force it through the round one. It asks what the new participant requires, then builds toward that answer even when the answer abandons the convenient path and the quick land grab that path promises.

Agents Differ From a Person at a Browser

The deepest reason for a dedicated transport is the caller’s nature. A human at a browser is present, reads the context, resolves ambiguity, and bears responsibility in person. An agent is a delegate. It carries authority granted elsewhere, operates at machine speed, composes actions across many parties, and leaves a trail someone will later need to verify with confidence.

A future filled with agents acting for people, for companies, and for one another requires a substrate that treats those properties as first-class citizens. Identity itself can prove. Authority scoped at the wire. Attribution that outlives the transaction. Replay safety is modeled on the networks that already move money safely every day. These belong in the foundation, available to every framework above, rather than being rebuilt differently inside each one and hoping to interoperate later.

A Substrate Rather Than a Rival

This is where the narrow waist matters most. The internet scaled because a thin, neutral layer carried everything above it, and everything below it plugged into that one shared point. AGTP aims at the same shape for agents. Frameworks such as MCP, A2A, and the rest become tenants of the substrate rather than rivals to it. The governance layer from earlier is a tenant too, drawing its identity and authority from the floor beneath it rather than rebuilding them from scratch. Each one gains a proven identity, attribution, and authority from below, rather than reinventing every primitive in a slightly different dialect that fails to compose. A reductive label misses this entirely, because the value lives in the composition rather than in any single feature an observer happens to notice first.

Designing for What Arrives Next

The strongest objection I hear reduces to one question, the same one an engineer put to me directly. Could HTTP do this with enough extensions? Possibly. Many things become possible with enough extensions piled high enough. The question worth answering is whether the result would serve agents as well as a substrate designed for them from the first line of the specification, with their reality as the starting point rather than a retrofit.

AGTP exists because the honest answer to that question is skepticism. Agents are arriving as a new class of participants on the network, distinct from a person clicking a web page, since that person was once from the paper mail they replaced. They deserve infrastructure shaped around their reality rather than handed down from ours out of habit. Calling AGTP a communication protocol, a semantics protocol, or an identity protocol, each captures a true fragment of the thing. The whole is a wager, and a deliberate one. The agentic era earns its own foundation, and not settling for what has already been built.

---

If you find this content valuable, please share it with your network.

Follow me for daily insights.

Book me to speak at your next event.

Start managing your agents for free.

Chris Hood is an AI strategist and author of the #1 Amazon Best Seller Infailible and Customer Transformation, and has been recognized as one of the Top 30 Global Gurus for Customer Experience. His latest book, Unmapping Customer Journeys, is available now!