232: The Agentic Web Is Still A Single-Player Game

[00:00:00] Slobodan: Amazon Web Services, Microsoft, Salesforce, SAP, Adobe, Bloomberg, Siemens, Deutsche Telekom. A total of 173 companies have publicly put their name to the same thing, a web standard, an emerging one called A2A, agent to agent, for AI agents belonging to different companies to find each other and talk to each other.

[00:00:21] Slobodan: So I looked at most of those companies. I took the web address, the project itself, A to A lists next to each name of the company, and I asked every one of them the question that the standard is built around.

[00:00:33] Slobodan: Are you there? What can you do? Only four of them answered in a way that the standard describes.

[00:00:40] Slobodan: This is No Hacks with me, Slobodan Manic. Nobody asked any of us if we wanted the internet changed. So check what AI is actually doing to it. The good, the bad, and the pointless

[00:00:54] Slobodan: Everything we get told about AI agents assumes that something is coming to us.

[00:00:59] Slobodan: There are agents that are going to interact with us, coming to our servers that can hold a conversation. Now, if that is true, I should be building my website to negotiate with this AI agent or swarms of agents or whatever. If that is not true, the real job is a lot more plain and a lot more boring, and I don't mind plain or boring, but a lot of what we are being told is, or it's sold actually, is for a visitor that has not turned up yet, including this A2A stuff.

[00:01:29] Slobodan: So I went looking.

[00:01:31] Slobodan: And the question for this week that this episode is going to try to answer is: Is there anybody on the other side of the fence just yet? And let me explain what I mean by that, there are many wrong answers to that question. I don't mean does anything respond. Th-there are responses.

[00:01:51] Slobodan: There, there are crawlers. There's ev- You, you can ask a tool for, I don't know, "What's the weather gonna be like?" And you get a response. It's that you get the weather back. You ask a website for a page, you get a page back. This happens on the internet all the time. What I mean here is, is there another party out there, someone with its own side of the deal, something that is trying to reason, something that can say no or come back with a different price or ask you a question you were not expecting, something that wants something that's not necessarily a human, another agent that my agent can talk to?

[00:02:24] Slobodan: Because this protocol, A2A, is all about agent-to-agent interactions. That is What agentic web is supposed to be like. Not automated crawlers that, that scrape a thousand or a million pages every day. N-not, not, not completely running that in isolation versus the rest of the world. I mean, connected and talking to each other.

[00:02:45] Slobodan: What I assumed going into this was that some of this was already happening in a small way. My agent and your agent, they could talk, they could loop us in when they need us, and they can sort something out either with our involvement or without it. What is actually happening, what is running out there is still a single player game.

[00:03:08] Slobodan: You send an agent out and everything it meets is, is kind of like scenery in a video game. Things that answer and just hand you over the answer you were looking for it's NPCs, if you ever played a game. Non-playable characters, the ones that the world provides for you to interact with while every other player is off in their own little copy of the world.

[00:03:30] Slobodan: Nothing wrong with, with games like this, but this is not what we are being promised with agentic web. And I want to be fair about what that does and doesn't mean. Now, the second player, the infrastructure for that second player is being built But what's not happening is anybody using it with somebody else, anywhere you or I can see.

[00:03:53] Slobodan: Agentic web is in that single player mode still, and the reason for that is not technical. The reason for that comes down to how many people have to agree with what's going to happen before anything starts happening.

[00:04:07] Slobodan: So let's go back to November 2024, trying to understand what's happening today. November 2024 was when Anthropic put something out that's called MCP. You heard of this. I'm sure you've heard of MCP. Now, here's what actually happens when you use an MCP server. You install a small program on your computer.

[00:04:32] Slobodan: Your assistant starts that program up when it needs it, and then your assistant talks to that server, that program. The assistant asks questions, "What can you do?" The program answers with a list, "Look up a customer," or, "Read this file," or, "Save this," or, "Upload. Check the calendar," whatever. And from then, your assistant can call any of those, get the answer back, the same way you would get through an API.

[00:04:56] Slobodan: This is just different interface for basically the same thing, and just carry on. There's no thinking, there's no reasoning on the other end. It, it's just returning-- It's responding to a command. That's the whole mechanism. It's two pieces of software on one side, on one machine talking to each other, and the-- that's it

[00:05:15] Slobodan: To start using that, for that to happen, for, for MCP to be interacted with, how many people do you think need to agree to any kind of deal? One. It, it's just you. It's just person connecting to an MCP server. So you take a tool, the MCP server that somebody else wrote, you wire it into your assistant, and you're just running it.

[00:05:38] Slobodan: The person who wrote it, they, they may have analytics to see what the usage is like, but they don't-- they, they're not in that conversation. There's no negotiation with that person. Nobody's signing on anything or anything like that. The specification says the client launches the server as a subprocess.

[00:05:56] Slobodan: So that server somewhere online that gives you some tools through an MCP server it's just running as a subprocess of whatever you're doing with your assistant or with your agent. So that means that the work starts on your machine because you initiated it, and then you're just waiting for a response.

[00:06:16] Slobodan: One person all alone just messing around with things online on the other da-end of that, you have a tool, you have the calendar, database, file, whatever, and it just answers. That tool just answers. It doesn't work for you. It doesn't work for anyone. It's just a simple tool,

[00:06:32] Slobodan: this is not a criticism of MCP. That tool is exactly what it says it's supposed to be. They're not lying. MCP is genuinely useful when implemented well. It can be great. But you can see what it does. And it, it, as the definition says, it's just a subprocess of whatever is running on your machine or inside your assistant.

[00:06:56] Slobodan: Now let's fast-forward to April twenty twenty-five. What is that? A year and a half ago, roughly. Google has put out A2A, Agent-to-Agent Protocol.

[00:07:10] Slobodan: This is different. This is not like MCP. This is not your assistant reaching for a tool no. Your company's agent, your agent, is going to talk to another company's agent Something you don't control, something you have never met, and it's not going to use tools.

[00:07:26] Slobodan: What happens there is going to be completely different. Your agent is going to go to that company's website. It picks up a small file, a JSON file that describes their agent. It's agent card.json, what they can handle, what the agent can handle, how to reach it, and then it sends over a task. So it's not calling a tool, it's sending a request.

[00:07:47] Slobodan: Not a command, a task that's written out the same way you would send it to someone you hire on Upwork. This is what I want. Then the other side takes that task, the other agent on that server works on it, and then when it's done, this could be in five seconds, this could be in two days, this could be in two weeks, a month, a year.

[00:08:06] Slobodan: It can last a very long time. It reports back on how it's going as it's working on it, and eventually it returns something or just refuses the request. So the far end of that, if you are the one initiating and, and just talking to another agent through that agent card, it's not a calendar or a tool. It, it is somebody else's machine running and somebody else's interest really on that machine with their own infrastructure from their own boss.

[00:08:35] Slobodan: It, it can say no to you because there's, as much as AI agents can reason, there's reasoning on that other end. It's not just a tool that you can use like a calculator or calendar or whatever else. This is the version of agentic web that when people think about what the agentic web is, this is the version everyone's been describing for two years.

[00:08:58] Slobodan: So same question. How many people need to agree to any kind of interaction before this starts? It's two, because two people, two organizations, your agent and their agent. There has to be a contract between them. There has to be a way to connect them, and there has to be somebody on both sides that's going to sign off to using this

[00:09:20] Slobodan: And that's kind of the gist of where the agentic web is right now. 'Cause MCP, MCP pays off on its own. You just wire it up. It can take an afternoon, it can take a weekend, it can take a week, whatever. And by that, you just, just plug it in, and then your assistant can do something with the MCP server that was not possible before you just plugged it on, turned it on.

[00:09:42] Slobodan: Agent to agent is very, very different. You can implement that perfectly, and you can get exactly nothing because the value does not arrive until somebody else you do not control has also done that same kind of work The first, we're going back in history, the first fax machine, to use that as an analogy.

[00:10:05] Slobodan: If there was only one fax machine in the world, that's as useless as it gets. Add another one, and they can interact. They can talk to each other. This is the difference between MCP that I'm sure you've heard of. I'm sure you, if you work for a company that has a product, digital product, I'm sure someone told you you need to implement MCP if you haven't already, or you tried them, whatever.

[00:10:29] Slobodan: That's the difference between MCP and A2A or agent-to-agent

[00:10:33] Slobodan: So one of them spreads at the speed of a person just, "I'm gonna mess around with this. I'm gonna try." The other one is completely different. It's two organizations. One request comes in, the other organization also needs to approve the request and decide what to do with it. It's not as simple as pushing a button to turn on the light.

[00:10:49] Slobodan: No, you push a button, and then it goes to the Ministry of Light, and they decided they want to do something with your request. Very, very different. And by the way, the launch of agent-to-agent in April 2025 was good.

[00:11:04] Slobodan: Let me explain how A2A works, because I think it matters. I think this is what has potential to define the future of the internet. Agent to agent, the protocol says an agent has to publish a small card, agentcard.json, describing itself.

[00:11:23] Slobodan: Uh, the whole thing, what it is, what it can do, how to reach it. And then there's a fold-- There's a path on your folder where that file needs to live. The same way you have robots.txt or that, that fake thing, lm.txt. The same way, there's a well-known/agentcard.json. It's always there. And, uh, every website has this app if, if it has it on the same address.

[00:11:52] Slobodan: You're not required to put it there, so this is not a test you can fail if you're running all those checks that Cloudflare has and that Google has in Lighthouse. If you don't have agent card because you don't have an agent operating on your website server, that, that doesn't mean you're failing. It just does-- It just means you're not using that certain technology.

[00:12:10] Slobodan: It is good. You don't have to worry about it at all

[00:12:14] Slobodan: . And if you look at this, that only four of these launch partners for this major protocol that was created by Google and then donated to Linux Foundation, only four of those launch partners have an open, accessible agentcard.json file.

[00:12:32] Slobodan: It tells you that either they don't care about this, and let's not jump to conclusions because I don't know, or they do care about this, but this is just not going to be on the open web. And I don't like that. I really don't So yeah, what that means for your server, for your website, and how this changes, how should you approach this?

[00:12:56] Slobodan: Well, the thing arriving at your server is not necessarily another company's agent that's negotiating with your company on a merger or whatever, like any silly thing like that. It's going to be a person's assistant, them running something in, in Claude or ChatGPT or Claude Code, any of those harnesses for, for, yeah, for LLMs, just alone doing stuff.

[00:13:24] Slobodan: And then that's who's visiting right now. That, that's who has been visiting the whole time. And that makes your website a scenery for that assistant. You do not get to do anything. You just participate as an NPC in their single-player video game. You do not make an offer or set terms or ask who's calling.

[00:13:44] Slobodan: This is me assuming that you do not have agent card and all this implemented, agent to agent implemented, because very few websites and very few companies do. So it's an assumption I dare to make. Now, preparing your website for this, it, it's directionally correct. It's the right instinct. It is the visitor that's being described to you that's, that, that's wrong.

[00:14:07] Slobodan: If an agent is just going to, you know, compare your offer to somebody else's and then come back with a counter, that, that, that's a real thing. That may arrive. It's not what's happening right now What's at your door on your website is somebody's assistant sent to fetch one specific thing, one page, compare pages with no ability to ask you a follow-up or try to talk to you even, and no patience for a page that's going to take two minutes for your agent to do something and then respond.

[00:14:34] Slobodan: That's just how it is right now. This is not, it's not negotiation. It's just, it's expecting an answer clearly, quickly, and yeah ' Cause most websites, if you think about it, really most websites, when anyone requests something, the answer is yes, unless they're blocking something or unless you're asking for s- for something that doesn't exist

[00:14:56] Slobodan: So we're still playing this as a single-player game. Doesn't have to be permanent, doesn't have to stay like that forever. I don't know if it will. What that would look like in, in, in this specific case is other people's agents just coming to your website and, and you being able to tell this is what happened.

[00:15:18] Slobodan: You cannot because analytics just doesn't exist. We're in the early days. So my message for you about this, you should, you should know what A2A is because it's a very simple standard. It's easy to figure out, it's easy to understand, and I have, I have blog posts about this on, on nohacks.co website.

[00:15:35] Slobodan: Forget about the vendor announcements. Forget about the launches, forget about the hoorah cham-champagne moments, and forget about the partners and the hundred and seventy-three company logos. Just watch your own logs, and if something turns up there that was clearly sent by another company system that's trying to do something, negotiate with you, there to complete a task, expecting a response from you that's not a simple server render of webpage, that's when this changes for you

[00:16:07] Slobodan: This is NoHacks. The notes and every source for this episode are at nohacks.co. I also send a newsletter. If you're interested in that, go to nohacks.co/subscribe, and please consider subscribing. My name is Slobodan Manic. Thank you for listening

[00:16:24] Slobodan: As promised, as promised, after we're done with the main topic, there's something that's been on my mind

[00:16:30] Slobodan: With everyone rushing to put AI into everything, to use AI for everything, mostly things that never needed to be done, let's be very honest about that, I keep wondering what that is going to do to our world. It was supposed to give us our time back a few years ago. That was kind of part of the promise.

[00:16:53] Slobodan: And then one of those CEOs, I'm not going to even say his name, said that people actually don't want to have free time a few months ago, a few weeks ago. What was supposed to happen was we are going to get the free time, but now we're more busy and everything is just shittier, and it's a race to the bottom.

[00:17:16] Slobodan: So what I wanna say is that-- I think rushing to implement more AI, to have more, more AI in your workflow, in whatever, I think that's a race to the bottom. I think that's the same thing as competing on prices for your services, and I think that leads to burnout and leads to nothing good.

[00:17:36] Slobodan: Not because AI is bad at this, but when everyone is trying to speed up and do more at the same time, and also more things that they've never done and they don't know how to have a good output for, nobody, nobody gets there earlier. Nobody gives you a better thing

[00:17:55] Slobodan: So don't focus on the output. This is something I've been saying all year. Output is free. I can create a 20-page document and send it to you in 20 seconds. I pray that you don't want to read it Focus on the friction. Add friction to everything you're doing because then you will care about it, and then you will have a connection to the things you're making and sending to others and sending out to the world.

[00:18:21] Slobodan: If you stayed this late and listened to the after episode, fully unscripted riff, you're a hero. Thank you so much. I'll talk to you next week