Learning Library

← Back to Library

Agent-to-Agent Protocols Redefine Software

Key Points

  • Google’s new Agent‑to‑Agent (A2A) protocols extend the recent Model Context Protocol (MCP) idea by enabling AI agents to discover, describe, and collaborate with each other, not just with tools.
  • For the past 70 years software has been built as deterministic, explicitly‑programmed logic, which limits flexibility because the system can only do exactly what developers code.
  • MCP cracks this paradigm by letting developers specify capabilities of tools in a structured way, allowing non‑deterministic LLM‑based agents to decide how to use those tools on the fly.
  • The combination of MCP and A2A represents a deeper shift: software’s underlying substrate moves from fixed instruction pipelines to dynamic, capability‑driven interactions between autonomous AI agents.

Full Transcript

# Agent-to-Agent Protocols Redefine Software **Source:** [https://www.youtube.com/watch?v=cPdVbVx5Z3Q](https://www.youtube.com/watch?v=cPdVbVx5Z3Q) **Duration:** 00:08:57 ## Summary - Google’s new Agent‑to‑Agent (A2A) protocols extend the recent Model Context Protocol (MCP) idea by enabling AI agents to discover, describe, and collaborate with each other, not just with tools. - For the past 70 years software has been built as deterministic, explicitly‑programmed logic, which limits flexibility because the system can only do exactly what developers code. - MCP cracks this paradigm by letting developers specify capabilities of tools in a structured way, allowing non‑deterministic LLM‑based agents to decide how to use those tools on the fly. - The combination of MCP and A2A represents a deeper shift: software’s underlying substrate moves from fixed instruction pipelines to dynamic, capability‑driven interactions between autonomous AI agents. ## Sections - [00:00:00](https://www.youtube.com/watch?v=cPdVbVx5Z3Q&t=0s) **MCP – Shifting AI Software Paradigm** - The speaker argues that Google’s agent‑to‑agent protocols and the Model Context Protocol represent a fundamental move away from deterministic, hand‑coded software toward AI‑driven architectures that can describe and autonomously use tools. - [00:03:05](https://www.youtube.com/watch?v=cPdVbVx5Z3Q&t=185s) **Challenges of Multi-Agent Coordination** - The speaker outlines engineering difficulties in building scalable multi-agent systems, highlighting state management consistency, reasoning overhead, and security risks. - [00:06:57](https://www.youtube.com/watch?v=cPdVbVx5Z3Q&t=417s) **Emergent Agent Collaboration Architecture** - The speaker outlines an open A2A standard that enables autonomous agents to discover capabilities, negotiate workflows, and form dynamic, observable software interactions—shifting from programmed code to intelligence‑driven collaboration while highlighting emerging security challenges. ## Full Transcript
0:00Today something really massive happened 0:02in AI architecture and I don't think 0:04most of us realize how big it is. Google 0:07announced agentto agent protocols or aa 0:10and everybody's focused on the 0:12impressive partner list. I really want 0:15to talk about something much more deep. 0:17What happens when we fundamentally 0:19change the substrate that software runs 0:21on? I've been obsessing over this lately 0:23because I've been deep in the weeds with 0:25model context protocol or MCP. I wrote 0:28an article on it on Substack yesterday, 0:30but fundamentally I think it represents 0:33a profound shift in how we build AI 0:36systems and the A2A announcement doubles 0:39down on that shift. So let me dive into 0:41it. So for the last call it 70 years, 0:44we've built software roughly the same 0:46way. Explicit instructions, explicit 0:49connections, explicit logic, what we 0:51would call deterministic software. Every 0:53interaction has to be programmed. Every 0:56data flow has to be defined. Every 1:00decision point needs to be all the way 1:02mapped out. This is how we got waterfall 1:04software, right? Uh and we've gotten 1:06really good at it. And it's 1:08fundamentally limiting. It limits 1:10because it constrains your upside. Your 1:13software can only do what you've told it 1:15to do. It can never do anything more. 1:17MCP was the first real crack in that 1:20paradigm. Because instead of programming 1:22exactly how an AI should use tools, MCP 1:26lets us talk about and describe tools 1:28and their capabilities in a structured 1:30way and then lets AI figure out how to 1:33use 1:34them. It's a subtle shift, but it's a 1:36really profound one. We're moving from 1:38explicitly programming to capability 1:40description. And what's interesting is 1:43that in and of itself is an evolution of 1:46the very concept of the large language 1:48model which itself is not deterministic. 1:51I mean people will argue with you and 1:53tell you that LLMs are inherently 1:54deterministic and that's a conversation 1:56for another day but from a usage 1:58perspective they don't function that 1:59way. Uh they function as 2:02non-deterministic software and we've all 2:04been getting used to that and that's why 2:05prompting is such a big deal. MCP really 2:08took that world and turned it into how 2:11we design software. And so now we have 2:16agents determining how they will use 2:18tools through MCP. And then it gets more 2:21interesting with today's announcement. 2:23And that's why this matters so much 2:24because A2A or agent to agent, it takes 2:28the same principle from MCP and applies 2:31it to how AI agents work with one 2:33another. It's not just about tools. It's 2:36about agents discovering each other. So, 2:39it's not just about agents discovering 2:40tools and using those tools and making 2:42tool allocation decisions. It's about 2:44agents discovering one another, 2:45understanding each other's capabilities, 2:47and figuring out how they 2:50collaborate. So, if you think about what 2:52that means, we're not just changing how 2:54our software works with tools and making 2:56that less deterministic. We're also 2:58changing how our software works with 3:01other software. 3:03So if you're an engineer, you're 3:05thinking about all the problems, right? 3:06How could this go wrong? You're totally 3:09right. Uh in fact, I have a lot of 3:10concerns with MCP as well. I think it's 3:13challenging to build scaled systems with 3:15MCP right now. Uh and I think that it's 3:18challenging to build with A2A right now. 3:20Doesn't mean it's impossible, but it's 3:21tough to outline a few of the issues. 3:24There's a state management problem when 3:26you have multiple agents working 3:27together, each with their own context, 3:29their own understanding. How do you 3:31maintain consistency? How do you prevent 3:33conflicts? How do you handle partial 3:34failures? They're not new problems per 3:37se. Distributed systems engineers have 3:40had to work on this before, but they 3:42take on new dimensions when your nodes 3:44are autonomous and when they your nodes 3:46make their own 3:47decisions. Second, there's reasoning 3:49overhead, which is an issue with MCP as 3:52well. Every time agents negotiate how to 3:54work together, they're burning compute, 3:56they're burning tokens, they're burning 3:57time. And in a multi- aent system, the 4:00cost will compound. So you need really 4:03sophisticated optimization strategies to 4:06keep the system efficient and 4:08performant. Unlike traditional software 4:10where we can optimize specific pathways, 4:13agent interactions are dynamic. They're 4:15unpredictable and it's more complex to 4:18optimize. Third, security remains a 4:22nightmare. I continue to just cry and 4:24pray for my friends who work in security 4:26because agentto agent interaction layers 4:28a whole new set of 4:29vulnerabilities. You need 4:31authentication, you need authorization, 4:33you need audit trails. Uh implementing 4:35all of this without destroying the 4:37flexibility that makes agent 4:38collaboration special. That's a 4:40non-trivial challenge. It's really 4:43tough. And what's what's interesting 4:46about this is that yes, these are all 4:48real challenges. They all will need to 4:51be addressed, but they're also inherent 4:53in what's being built here. They're also 4:55features in a sense. We are being forced 4:57to rethink fundamental assumptions about 4:59how software should 5:01work. In traditional software, we would 5:05optimize for that predictability. Every 5:07pathway would be known. Interactions 5:09would be defined. All of the outcomes 5:12would be mapped out. We would build 5:15Salesforce. I kid. I kid. Uh but in an 5:18agent-based 5:19system, we have to optimize for 5:22adaptability and flexibility. That's 5:24kind of the point. You want to optimize 5:26for moving intelligence from the human 5:28layer down to the agent layer. We aren't 5:32going to be able to predict every 5:33interaction in that world. We're going 5:35to have to be able to build systems that 5:37can handle 5:39emergence. And that's a really profound 5:41shift. And this is why the combination 5:44of model context protocol and A2A is so 5:46profoundly powerful. MCP gives agents 5:50the ability to understand and to use 5:52tools. A2A gives them the ability to 5:55work together. When you put that all all 5:59together, when you combine it, you're 6:00creating the foundation for truly 6:03autonomous software systems. 6:08So if we go to a concrete example, let's 6:10say you're building a system that helps 6:11with call it sales ops. In a traditional 6:14architecture, you'd have to explicitly 6:16program all the workflows. You'd have to 6:18qualify the leads, when to follow up, 6:20how to route different types of 6:21inquiries. It would all have to be hard 6:24hardcoded. Let's layer on some of that 6:26agent complexity now and see what 6:28happens. With MCP, you can give an agent 6:30access to your CRM tools, your email 6:32systems, and your analytic platforms. It 6:34then figures out how to use those tools 6:36to accomplish the task. Now, with A2A, 6:39something more magical happens. Your 6:41sales agent can discover and collaborate 6:44with other specialized agents. Maybe 6:47there's an agent that's really good at 6:48writing email copy, or another that's 6:51expert at pricing analysis, another that 6:53specializes in calendar scheduling. 6:55Instead of programming how those 6:57capabilities 6:58interact, you let the agents figure it 7:01out. They discover each other's 7:02capabilities and negotiate on 7:04collaboration and dynamically form 7:06workflows based on the specific needs of 7:09the 7:10situation. If you look at A2A specs, you 7:13can see how carefully it's been designed 7:15to make this kind of collaboration 7:17practical. It builds on existing 7:19standards like HTTP, JSON RPC. It 7:22supports longunning tasks and different 7:25modalities. It's designed to be 7:27observable, which you definitely need. 7:29It's designed to be debugable. 7:31Uh, and critically, I think especially 7:33if it's going to work with MCP and kind 7:34of the movement there. It's designed to 7:36be open. This isn't just another 7:37proprietary standard, it's supposed to 7:39be an invitation to reimagine a Gentic 7:42software. 7:45I think what we're watching in real time 7:46is the emergence of a new kind of 7:48software architecture, a new software 7:50paradigm. One where the boundaries 7:53between systems become much more 7:55permeable which introduces, as I called 7:57out, lots of security issues, but it 8:00also introduces a lot of 8:02opportunities. Functionality isn't just 8:04called, it's negotiated. Now, 8:06intelligence isn't just a feature, it's 8:08sort of in the fundamental substrate. We 8:11are delegating to intelligence instead 8:13of delegating to software. And that's a 8:15fundamental shift. Yes, there's big 8:17challenges ahead. Yes, we're going to 8:19discover whole new classes of problems, 8:21whole new classes of frankly headline 8:24and defining 8:25issues. Uh, and that's okay. That's 8:29exactly what makes this exciting. We're 8:30not just adding new features to the same 8:32old software. We're fundamentally 8:34rethinking what software can be. 8:37And so to me, like if you think about a 8:38world where we can combine MCP, we 8:40combine 8:41A2A, this is the beginning 8:44of truly automated software systems. And 8:48I think that's going to be a massive 8:50deal. I think that's going to change 8:51everything. 8:54What do you think?