Agentic AI and A2A in 2025: From Prompts to Processes  ## Metadata - Author: [[The New Stack Podcast]] - Full Title: Agentic AI and A2A in 2025: From Prompts to Processes - Category: #podcasts - URL: https://share.snipd.com/episode/1bbaa9b0-e615-4e55-a270-8c1b55f63b17 ## Document Notes ## Highlights - **Agentic AI Enables Autonomous Actions** - Agentic AI moves beyond chatbots by reasoning, summarizing, and autonomously driving actions in business processes. - This shift enables enterprises to gain real value by deploying AI agents to tasks like order processing and banking transactions. Transcript: Kevin Laughridge Know, it's interesting. Agents is the new buzzword. A year ago, we were talking Gen AI. And I think they're similar. But the differentiating factor on an agent is really, can it do something in which it's pulling information, it's summarizing it, it's doing the reasoning associated with it, and Then we're actually driving an action. I'm going to summarize. So if I think about some of the technology out there today, like a chat bot, I ask it a question, it gives me an answer. I need it to do something to really be an agent, whether that's send an email, whether that's actually process an invoice. And that's where Google is starting to go, which is they understand for AI to be really useful, especially in like the enterprise software market. We need these agents to do things so we can get value out of it. And then we can deploy those things into business processes, whether they're things like order to cash or whether we're actually taking financial services and banking transactions And processing them with agents. ([Time 0:01:06](https://share.snipd.com/snip/2aa925ba-c562-43c4-ac4b-38bcf4ed0845)) - **A2A Protocol Connects AI Agents** - Google's open Agent2Agent (A2A) protocol connects AI agents across vendors to collaborate in enterprise ecosystems. - Stitching diverse AI solutions creates optimized business processes beyond single-agent capabilities. Transcript: Kevin Laughridge I'll go back to the keynote where they talked about the agent to agent protocol. We're extremely excited about that because it's an open protocol that allows us to connect multiple ISVs together, multiple different software companies so that we can actually Stitch together business processes. One of the things that we see, and so again, Deloitte's business is very much supporting large industries to solve problems. You can't be monolithic. We have to play within an entire ecosystem, and we need agents that actually communicate and do things together. Because we never have one agent or one solution that solves the entire business process. We call it a string of pearls. I need lots of different solutions. Sometimes they're robotics. Sometimes I have actually data analytics and data modernization I need to do. ([Time 0:04:10](https://share.snipd.com/snip/dcf7c910-0093-480e-9572-819221f698a5)) - **A2A Integrates with Existing IT** - Enterprises cannot rip and replace decades of IT; A2A fits into existing systems to enable AI-driven collaboration. - Adoption of an open protocol by major vendors accelerates AI integration and enterprise transformation. Transcript: Alex Williams You talked about the ADA protocol. Why is that important to you? What is it significant about it? Kevin Laughridge Companies have been built in their IT tooling for decades. Ripping and replacing everything is impossible. We need to fit within an ecosystem that is mature, that is stable, and that works. Agent-to allows us to connect into existing solutions without having to replumb the entire enterprise. I'm really excited because we're going to have to change how companies interact and work, and we see AI that will help drive efficiency and speed, but we need to be able to work within An existing set of technology. Having an open protocol like A2A allows us to do that. Having it supported by Google, who is an industry leader in terms of AI and actually can drive this change and other ISVs jump on that same protocol, I think is really important. ([Time 0:05:07](https://share.snipd.com/snip/ada7a2f7-17b8-434c-ad0b-8bc9636c9e85)) - **Build AI Platforms for Scale** - Organizations must develop AI platforms rather than individual agents to achieve production-scale AI deployments. - Building a 'model garden' and platform delivers efficiency, scalability, and integration potential for enterprise AI. Transcript: Kevin Laughridge Of our clients that were first to market. And it was go through and say, I want to go and investigate, are these use cases, can I actually get value out of these use cases? And so in 2023 and 2024, it was heavy. What can I do with it? And can I experiment with it? And they started pulling it in and then started to say, can I build this into my enterprise architecture? 2024, we've seen we're actually getting them into production and they're starting to consume them. Not at the scale I think that originally had thought. And so now 2025 is, again, we believe that's the year where we're going to see productionalized scale of this. What they're finding and my clients are finding is they need optionality. It's not one LLM solves all the problems. I need more of a model garden and I need to be able to swap different LLMs based upon the type of problem that I have. It's also, we started very much of individual use case by use case. And the reality is, is if you just build by use case, it gets pretty expensive. And so that's turned to, I need an AI platform and I need to actually have a place that all of my agents can run. And the more agents I build on that, I actually get efficiencies of scale. And so we're pivoting more towards, I need to build a platform for that. Kind of like decades ago when we decided we needed an enterprise data warehouse. I'd originally had lots of different databases sitting around the enterprise. We started to consolidate that so I could manage my data in a very clear way. ([Time 0:08:50](https://share.snipd.com/snip/eda44ae8-329c-478b-af98-b88c422e2d5e)) - **Agentic Frameworks as Platform Components** - Agentic frameworks are integral parts of AI platforms that include UI, runtimes, and connectors. - Connectors enable agents to interface with multiple business systems efficiently within a unified platform. Transcript: Kevin Laughridge Should end up being part of the platform. So my overall agentic framework in terms of I need a specific sets of capabilities, whether I need the actual UI that's part of it, I need a runtime framework. I need to actually have connectors. I view that as all part of that AI platform. And that's where we're going to get the value. Because if I have to plumb individual agents to every single one of my systems one off, it gets really expensive and it's hard to actually maintain and manage. And so as I get to this platform and my connectors, and I'll go back to Google's agent space, of the really big powerful things I see in agent space is all the connectors. Because once I've plumbed it into each of my different systems, I can now turn on agents and use that plumbing to do things with it. ([Time 0:10:16](https://share.snipd.com/snip/763bb937-eebc-4080-ac8a-1066d1362081)) - **Manage AI Plumbing Complexity Centrally** - Manage plumbing complexity by centralizing connectors and governance within an AI platform. - Use telemetry and API cataloging methods to monitor and maintain AI agent interactions and integrations. Transcript: Alex Williams So, you know, one of the biggest complexities we see is in configurations, right? And, you know, I'm curious on like how that complexity is gonna be managed over time. Like, how do you see what you're doing now as addressing that complexity issue? Kevin Laughridge At the end of the day, connecting applications 20 years ago was complex, and today it's still complex. And so that plumbing will end up being there. It will still end up being complex. If I at least have my AI platform and I met connectors that I'm focused on managing and governing within that platform, I have better visibility to each of these different connections. But we had this similar problem when we were talking APIs just a few years ago when I needed to have an API catalog and I needed to actually monitor and manage that. I had telemetry. That's where things like Apigee came into play. I think it's a very similar type of thing that we'll have to manage with AI because ultimately we need to have applications of these agents talk to each other. ([Time 0:11:07](https://share.snipd.com/snip/bf787189-37ed-4c1f-8909-f5f75b042702)) - **A2A Enables Future Autonomous Agents** - A2A is a protocol defining how autonomous agents communicate within frameworks. - Human oversight remains vital due to risk, but autonomous agent adoption will grow in the next year. Transcript: Kevin Laughridge See that overall our ADA ends up being a protocol. Framework says, I need this type of a capability. The protocol is actually the way that they substantiated that capability. Alex Williams Does that create then the capability for autonomous agents? It should. Where agents talk to each other. Yes. Have you seen that yet? Kevin Laughridge We're getting there. We still have human in the loop in the majority of our agents that are customer facing or they're not customer facing. Like at the end of the day, most of our clients don't want to end up in the wall street. We've seen a couple of missteps years, you know, that have happened. It's a, I don't have, I the risk associated with turning on an agent yet. So I keep a human in the loop to make sure that it's an ability for the human to say, I don't agree with that and that solution. That's not to say we won't get there. And I think in the next 12 months, we're going to see more and more autonomous agents. ([Time 0:12:14](https://share.snipd.com/snip/ce09c3b0-8e14-4c31-8bba-6f2d4e96ec35)) - **AI Resolves Legacy Code Debt** - Deloitte helped a client with 300 undocumented Pascal applications by using AI to understand and generate requirements. - This accelerates technical debt resolution much faster than traditional methods requiring specialized programmers. Transcript: Kevin Laughridge It needs to be managed, and it's going to be there. But I'll say one of the things that we're finding with AI is we're actually able to work through technical debt, prior technical debt, faster. So in fact, I just got a call from a client. They have 300 applications that have absolutely no documentation. It's written on Pascal and they don't understand what the application does. And we're using AI to actually unpack that and start to build the requirements directly off of old code. That's pretty interesting of how we could advance it, because before we used to have to hire Pascal programmers, if we didn't have them on staff, and actually go through and document And code it. ([Time 0:13:45](https://share.snipd.com/snip/1952b97f-6b1f-4ce8-a9fe-7df91616ad47)) - **Use AI as Developer Pair Programming** - Leverage AI agents as active pairs in software development to automate breaking down requirements into stories and test cases. - This pairing improves productivity and quality by linking requirements directly to coding and testing. Transcript: Kevin Laughridge I'll give you my perspective and answer, and maybe it's, I think it's aligned on pairing. We've built a couple of AI-enabled applications that really look at an entire business process, and we call it AI Assist. And again, it's a part of, it's not using the agent to go autonomously and do things on its own. It's actually the engineer and the product owner are using these agents at different points in the overall software development lifecycle process to help accelerate and do things That, candidly, we should be doing anyway. So as an example, we'll start with a business requirement. We'll use AI to actually break that business requirement down into stories. And then those stories, we use AI to actually groom, it goes to the backlog from what we've done for the last five years. And the AI will tune the stories based upon history. So it'll find how we've done logging before and actually build those pieces in. And then it will take those stories. The human has an ability to change it because maybe it doesn't come out perfect. Then we can take that and actually start building test cases, not just positive but also negative. And we always should be doing both positive and negative test cases. But in every one of my projects, that's a place where sometimes you may not necessarily do it as much. And so the pairing concept, what we see it is it helps drive more efficiency to actually drive delivery of the process as designed. And we're getting better, higher quality output because we're doing more testing. We have more information related to the product stories. And one of the things we built on this, which I used to be a programmer decades ago and I never had it, but we'd have the requirement. You never actually had a connection from the requirements repository into the IDE to code And we're actually making that connection now. And then we get into the IDE and we use AI to actually break that requirement down into a bunch of steps, almost pseudocode-like, and then we start coding that. So that concept of AI accelerating, we still have engineers and our existing staff, but it's giving them the ability to be more efficient. ([Time 0:16:36](https://share.snipd.com/snip/1ff3b0f5-c390-45ed-90da-dad5be137660))