Enterprise AI Agents and Multi-Agentic Systems with Tredence and Google Cloud: From Concept to Production

Speaker A: Hey everyone, and thank you for joining this special edition of Treatence's Beyond Possible Dialogues. My name is Alex and I'll be your host for the next 30 minutes. Joining me today are three leaders and close friends that I have the privilege of working with, Unmesh, Sala and Maruti. Over the next few minutes, we're going to go and talk about how organizations are looking at agentic AI and moving from concept to production. We're gonna cover some of the common pitfalls and some practical examples of how to get started. Let's kick off with a quick overview of who's joining me today. First is Maruti, a tenured Googler in the partner engineering space, specifically focused on agentic AI and helping customers and partners adopt Google's latest technology. Also we have Sala, uh, another leader within Google's partner organization, specializing on how partners and customers can best work together to unlock tangible business value. And last but certainly not least, is Enme, the SVP of AI at uh Treatence, who has decades of experience solving data and AI problems for the world's largest organization. With all the hype and pace of innovation, I'm sure many people feel like they need an agent to just understand where to start. Luckily, today we have Enmeshed, Sala and Maruti to unpack this topic. Let's go ahead and dive in.

Speaker B: Thanks for having me here. Super excited to be on this podcast today to talk about how Credence and Google Cloud are helping our different kind of customers with their different various complex challenges. And also excited to be on the same room with my friends Unmesh and uh, Sala, where we talk about AI agents day in, day out.

Speaker C: Hey everyone. Hey Alex. Unmesh, Maruti, thank you very much and really happy to join the show. The work you guys are doing to shed lights on these most complex topics is fantastic, frankly and I'm honored to be part of this conversation today. Thank you.

Speaker D: Hi everybody. Great to be here and thanks for inviting me here. I'm Anvesh Kulkarni. I'm the Senior Vice President of AI at uh Credence, a uh, specialist services provider in the field of data and AI. In my role, I work very closely with our customers on their AI implementations and I also work very closely with our partners like Google. So I have a very vantage point and I'm very happy to be here to share it and also learn from our partners here.

Speaker A: Thank you. Awesome. So, so over the next 30 minutes we're going to really unpack how AI agents and multi agent systems are really disrupting how businesses do business. With all the hype and the pace of innovation, I'm sure many people feel like they need probably an agent to understand, you know, where to start and how even to interpret everything going on. Lucky for us, today we've got Unmesh, Sala and Maruti to unpack this topic. Let's go ahead and dive on in Unmesh, Maruti, Sala, you all sit at the intersection of technology and business transformation together. Unmesh and Maruti, you just co authored an article on how to take AI agents and multi agent systems from concept to production. And we've all heard about our agents actually making it in production. Are they actually delivering an impact? What I'd love to start with is from each of you, is there a specific experience or challenge that helped you or drove you to publish this blog and Mesh? Would you mind kicking it off for the group?

Speaker D: That's a really great question, Alex. Uh, and I would say that the way this article, the blog was born was very organic, right? As part of our, uh, the strategy and implementation work that we do with our customers, I get to interact with a lot of very senior AI data and business leaders. And I clearly see that there is a widespread excitement about the evolution of AI agents and application of agents. But there was also a, uh, need for clarity. There's confusion about how to take the pilots, the POCs that people have developed over the last years and take them to production. There were questions about, you know, ROI governments. There are questions about common themes like what are these agent big systems, right? How should I get started with it or should be my right platform? So we were getting these questions and with Maruti and Sala and others, we were constantly kind of responding and addressing these concerns. So we thought, you know, why not just list out the most common questions we get and address them through a, uh, series of blogs. So that's exactly why the blog happened. Now I can just point everybody to the blog and we can move on to the business outcomes part. That's the genesis of Blocky.

Speaker A: So Maruti, based on everything Unmesh just said, could you talk a little bit about the pace of innovation within Google Cloud that kind of helped drive the blog on your end?

Speaker B: Yeah, sure. Like Unmesh mentioned, I think the core challenge or the main challenge enterprises are facing is to keep up with the latest advancements that are happening, adapt to, and ultimately how they can take advantage of these, uh, technologies and products to build uh, the new AI agents. Right now on one hand, companies like Google are coming up with latest uh, advancements starting from the model versions, the way the Gemini models or flagship models have evolved over the time and also the tools to actually build agents from unified platform starting from application level utilities or tools or products and also the infrastructure level Google is advancing or uh, building this at a rapid phase, these production tools. Now on the other side, partners like Tridents have actually helping handhold our customers to take their use cases to production. And now if you see it's kind of evident and crucial for both of us like Google Cloud and Tridents to partner and collaborate and come up with these kind of sharing our experience from the field and even help people understand, you know, clear the noise uh, and get the signals in terms of how should they build AI, ah, agents and take them from like Unmesh mentioned pilot or POC or Idea 2 production and that kind of the main motivation for us to come up with uh, not just one blog, we just started with one blog but the idea is to come up with a series of blogs. So uh, keep a tab on uh, the blog pages that um, tradence, uh, um, has the website Sal, what would you add here?

Speaker A: As we kind of look at that journey companies are going on on the vision of concept to production.

Speaker C: Great points from Earth and Illumination. One thing I'd echo is it's not necessarily always technology. That's the short end of the stick. It's often perfection is the enemy of progress rather than mulling and having a perfect plan, even in the face of agentic AI, taking that first step early is critical. So we're seeing an incredible pace of innovation within Google Cloud and with that really comes a responsibility to ensure our customers can actually harness these capabilities to drive real business outcomes. So that's really where our superpowers lie in sort of shortcutting that process and where our partners become like yourselves, become absolutely crucial. They're the ones really who are doing all the work in the ground, working hand in hand with customers to translate the power of technology into tangible business outcomes to solve their specific and unique problems and challenges. So as the two gentlemen um, co authored the article, it really was born from the need to provide a more clear roadmap for our partners as well as customers to navigate this exciting new landscape of agentic AI and successfully taking these powerful concepts to production.

Speaker A: I couldn't agree more. And I think if we go back to where Unmesh started with even some of the confusion that exists in the market and the purpose of that blog, Maruti, one of the questions I Think you might be really well positioned to answer is with all of this confusion on agents and then traditional software and genai, can you really describe to our listeners what exactly is an AI agent? How should they be thinking about it?

Speaker B: Every conversation or every blog starts trying to address this question. What is an AI agent? I think um, because um, every organization, every enterprise has kind of a same foundation of the definition but also they try to adapt or define in a way which is more suitable for their needs. Right, so, but again very high level zooming out from a Google point of view or after talking to a lot of partners and customers, I think we try to keep it very simple. Right. Think of an AI agent is kind of a software application that people are used to writing or building. But again here you're relying on a large language model model or a reasoning layer to actually decide the flow of the application. So in a way you're actually letting the software application take actions on behalf of the user. So basically in a technical way if you say the software program uses the LLM or the reasoning layer within Google, or we call it as a brain, which actually reasons and then builds a plan. And now again if you talk about reasoning, which is more like uh, you know, if a user sends a request to AI agent, it actually takes that and kind of distills or breaks down into subtasks or do like chain of thought of prompting different algorithms and then it kind of builds a plan and then agent kind of executes that by taking advantage of what we call tools now tools. Um, again think of a uh, way for these LLMs or agents to talk to external world whether it's maybe APIs or uh, databases or even enterprise systems. And I think that is where even enterprises are struggling. And that is where I think Google is building a lot of products to simplify these things. Like our uh, adk, the Agent Development Kit, which is an open source framework to simplify the low level tasks or aspects of building AI agents and make it very easy. And also the Gemini Enterprise which we kind of launched just yesterday, uh, which kind of heavily focus on how to bring your enterprise systems very close to your AI interfaces or your AI agent. That way you don't need to worry about the underlying challenges or aspects of managing those M connections or integrations, et cetera. And then third, I think end of the day these AI agents are to augment humans. So now end of the day I think you need to figure out how to make your existing applications agentic in nature and also advance or enhance the way Users usually typically interact with the agents.

Speaker A: That's a great point. And I think there's some remarkable capabilities coming out from Google Cloud. And many organizations are trying to figure out where do they get started, how do they leverage agentic AI? And you know, is it a agent, is it multiple agent? And I think the reality is it can be sometimes hard to even understand how our organization is doing this successfully. Unmesh, could you give an example for our group about a practical real world example of a multi agent system and what it looks like and you know, how a business is using it.

Speaker D: Wow, an example. And that's um, it's so hard because there are so many really good examples we have seen in the last 12 to 18 months. Uh, let me pick one that has recently won an industry award for its architecture and the outcomes it delivered to the organization. Here we worked with a CDO and CMO of a large CPG company. It used to take them about three months to launch a campaign. Multiple regions, multiple languages, very complex businesses. And with Agentix solution, we deployed agents. One agent actually analyzing the past campaign performance, understanding what resonates well with a particular audience type another agent that will generate new content dynamically so it can just be refined by the creative teams. But a lot of early iterations could be done with just issuing prompts to AI models. The third one, which is very interesting, is an agent that actually checked if the generated marketing content is actually compliant with the company's policy, is it on brac, those kind of things. And finally an agent that actually helps you launch and integrate with the marketing resource management processes and the backend systems like that. So now you look at an agentic solution which if you are a marketing manager now, instead of launching a campaign in three months, you're launching four different campaigns. You're taking content for a major market and with models like Google's new Nano Banana, you can quickly use it to create content for a smaller regional market, maybe a local language. So massive roi, you've seen that the company was able to shrink its campaign launch time. The customer experience metrics have gone up and they've obviously been able to reduce the cost of sandwich. So that's one of my favorite examples of how agents come together to solve a business problem.

Speaker A: That's a wonderful example. And mesh. And one of my favorite pieces of that example is a lot of times there's a lot of companies who are saying, oh no, AI agents, how are they going to replace our workforce? This is a great scenario where they're augmenting the workforce, they are taking something and being able to do it at a scale that was previously never possible, never cost effective to run these campaigns at scale and enable this level of personalization and this response time using their existing team in place. Now as we look at uses of agentic AI and multi agent systems, Sala, you have a really unique perspective kind of based on where you sit within Google Cloud and the partners and the customers you interact with. Would love to get your point of view on how you're seeing agents and multi agent systems get traction from where you've said.

Speaker C: Great question Alex and um, Mesh, there was a very inspiring example in cpg we often talk about whether the right examples and there is no such thing as a right example, the ones that deliver value are the right examples. So I think you touched on that. So Google has a really profound culture of dogfooding and just relentless self improvement. One of our 10 company philosophies is that great just isn't good enough. Which is why we keep disrupting ourselves at Google since the earliest Palm models to our latest Gemini models, nanopanas of the world, we've just been iterating course correcting and repeating that virtuous cycle with customers, partners like yourselves and analysts. So that said, every technology and business solution that we offer it has a generative and AI component baked into the heart of the solution. We've gone really past that single agent experience we were probably experimenting with uh, 12 months ago, not far off to very complex multi agent workflows, orchestrations and so forth. So we're integrating at Google we're integrating and directly into all of our products to enhance developer uh, experience, to enhance our data science experience. Think of uh, tools like codesys for instance which really acts as an AI powered collaborator for developers or data agents that really streamlines things like data engineering and pipeline science workflows. Let's not forget what Mariti talked about earlier. Um, yesterday we announced uh, Gemini Enterprise which is really a foundational shift in positioning our agentic platform as that front door for AI in the workplace. Which really brings me to the next point which is that it's not about the agent, it's the workflows. Rethinking and reimagining entire workflows and how sops and businesses are done. So for example we talked about nanobanana earlier. The amount of work and acceleration we've seen around creatives, marketing and personalizing presentations say for our customers using similar Gemini models. It's very real, it's happening. But none of these are like disparate tools. Thomas shared a anecdote. Yesterday was like 246 interactions a day. I think that's at the low end, I believe, because the customers I'm speaking with, they're just like, we're extensively using it and we're accelerating the demand and the outcomes and the outputs of it. No one is talking about hallucinations anymore. There is a reason why. And, um, that's because the controls and the guardrails we have in place go far beyond the hallucination challenge we had about 12, 18 months m ago. But again, none of these are like disparate tools. The outcomes are exponentially better when multiple agents and tools come together as part of a business workflow. This is something that was never possible before at the scale of a business like Google Cloud. It's this rewiring that we're talking about, you know, uh, our, uh, products and internal processes that's really transformative.

Speaker A: Absolutely. It is remarkable to see what you all are doing and it's really as a partner of Google Cloud, seeing how you all leverage it internally and then also how that translates to the value customers are seeing is just fascinating. And it's changing every single day now. I think for this group here, one of the things as we talk about change and the pace of innovation is there's a lot of pitfalls. And I think there's also a lot of lessons learned over this time. I'd love to maybe start and maybe a few of you can chime in, but unmesh, maybe you kick it off for us as all of the opportunities with this. It almost seems endless. What are some of these lessons learned or pitfalls that you've personally seen or heard about that organizations should be keeping an eye out for?

Speaker D: That's another great question because, you know, surveys and analyst reports have repeatedly pointed out that, uh, a huge fraction of the POCs and pilots actually don't make it to the production or they don't deliver the ROI that we expect. Right now our experience has been somewhat different and I think it's important to kind of glean some of these lessons. Learn what is making these pilots not go to production. Part of it is just, you know, people are trying so many different things that they are picking up things to actually take to production. So the numbers are a bit skewed. But generally speaking, AI leaders tend to overestimate the impact of agent. Now it's such a big buzzword that everything is getting agentified and we to step back and be very choosy. About where would you use a very semi autonomous intelligent system and make sure that the rest of the organization comes along, that the people are ready to do it and you have the resources and the skill set and the training and everything. But part of it is, and this MIT study actually points it out, that organizations who tend to kind of over promise or go and try to do it all by themselves are not as successful. So making sure that you have the right platform, the right clouds, the right implementation partner who's done it in many places, bringing them along as advisors, as implementation partners is actually one critical lesson or critical ingredient, uh, to success. The second one is actually very, very common for any new technology. When you are moving from pilot to production, you absolutely have to meet the need for governance and explainability. And this is especially true for something as semi autonomous or autonomous as agents. So making sure that that's built in, as Sala mentioned earlier, the guardrails, the explainability of agents, the idea that you can ground it to truths. And there is obviously a human in the loop, especially early on as you build and roll out these systems. Those would be key lessons that I think I have gleaned through our implementations over the last couple of years. But I'm pretty sure Maruti Sala, you have seen many other implementations. So what's your take on these things?

Speaker B: Those are very good, uh, considerations you called out there. I think from a technical standpoint I would want to add like three pitfalls or challenges our enterprise or customers are running into. Right. One, because these are little different than the traditional applications. And also as this autonomy or the way AI agents run is changing or uh, evolving, I think, uh, enterprise are struggling to figure out how do they manage the lifecycle of the AI agents, which is again you can call it as AI ops for AI agents. I think, which again I think Unmesh and Sala talked about how do you put like guardrails, governance? I think now it all boils down to how do you put like your agent operations, uh, starting from how do you build, deploy, uh, monitor and also evaluate. I think evaluate is one of the biggest aspects or biggest challenges as uh, people build AI agents. The second one is integration challenges. Now AI agents are not powerful without actually taking advantage of the integrations, talking to their enterprise systems databases and other agents as well. Now I think this is where people are struggling and Sala alluded to, and I mentioned that Gemma Enterprise was launched yesterday. It kind of provides a managed way of interacting with uh, commonly used or highly used enterprise systems within our organization. And third, again, this is not new, but there's security and uh, uh, safety aspects of building applications. Right. I think Sala talked about guardrails. Again, I think putting the right guardrails in terms of what agents can access versus not, I think the key as enterprises design their AI agents. And also the other thing is from a responsibility or explainability point of view, how can you put safety measures so the AI agents are not executing or doing things that are harmful or not intended? I think these are some of the technical challenges or pitfalls we are seeing, um, as enterprises build their AI agents.

Speaker C: You shouldn't be building an agent just for the sake of having an agent. Nor are agents a default response to everything. Uh, it generates a lot of excitement. Almost every conversation I am in with CXOs and board members and like we want to do this agent, yet their data journey is not quite mature enough to actually drive outcomes with those through AI. Uh, the most important one is to really ensure that you as a team have executive buying. That's by far the most important thing. With a healthy dose of their, their expectations being grounded in reality. Because AI is not a panacea. Selling a business case on starting small and doing one thing really well, exceptionally well, learning from IT and scaling to additional use cases. Which is why I'd rather say that I'm fully grounded on the people element of the problem more than the technology side of it. Technology is great. It's evolving at a rate that people like myself can absolutely not keep up with it yet. This is another reason why our partners are so crucial in this ecosystem. They're the ones who are bridging that gap and accelerating value. We rely on our partners to just help us scale the expertise needed to implement these complex solutions safely and effectively. Without the right skills and the right partners, even the most promising agentic AI projects can store and mesh.

Speaker A: Maru Tisala, thank you so much for the quick recap around some of the challenges and pain points and lessons learned um, over the last few months, years and even just weeks or days in agentic AI innovation. The pace we're seeing is rapid and the insights you've provided over the last several minutes have been really helpful in understanding where is the industry going and how should organizations be thinking about this Now I'm sure each of you could spend also days, weeks talking to organizations and advising them on how to be going forward, but we've only got a few minutes left. I'd love if each of you could maybe chime in for our listeners with, you know what's a single piece of advice you'd give our listeners to prepare their organizations for this next era of agentic AI.

Speaker D: Let me jump in. So you know, at Credence we see different companies, especially enterprises, at different points along the spectrum of agent TK implementation. And some companies are actually taking a very cautious approach and like let me wait and see how it goes and kind of stop at pilots. And my advice to these organizational leaders is to actually take measured risks here. So understand the benefits and costs of agentic approach, bet on the right cloud stacks, bet on the right implementation partners, but don't stop because the risk of waiting is also too high. If you don't embrace this change, you'll get obsolete and will likely fall behind forever. Right. So it's important to pick the right partners as far as implementation goes. It's actually useful to have a very good checklist of things like what are the core considerations, what are the core platform choices, how do I go about it? And on that point I would recommend everybody to actually start with our blog that's on tridents.com website and actually use that as one of your checklist building tools. And of course feel free to reach out to us at Google or at Credence if you are looking for some very concrete advice on taking your pilots and POCs to production scale.

Speaker A: Murthy, what would you add?

Speaker B: I think you asked for advice, right? I think uh, I tell this to a lot of uh, partners and customers these days is uh, even though these are new way of uh, building applications, the core concepts or core foundations that are required for any software application like security, uh, like the team talked about in the last few minutes, Security, how do you monitor these applications and how do you manage the lifecycle of these applications or AI agents is very crucial. I think as an organization you should have a foundation in terms of the definition of AI agents. Starting with the definition of AI agents and um, the purpose, the scope as you build before you build, like Sala mentioned, just don't build for the sake of building. So you need to have like a foundation philosophy as a guide blueprint, uh, I don't know what you call that, but a system in place which kind of talks about how these aspects which are very crucial or foundation for building any application applies to AI agents. I think then you can talk about how you can apply to different domains, different use cases or you know, different customer scenarios.

Speaker C: That's a great response. Maruti and Nonmesh. And as you guys were talking it just dawned upon me, uh, one of My mentors in my early days during when cloud native was a thing yet said something along the lines of find a workload, run a workload. And it applies regardless whether it's the AI era. We may have AGI down the road in a couple of months, a year. I think it's very applicable. My one piece of advice would be to start small but think big. Absolutely. You have to have a grand vision. Just don't try to boil the ocean. Identify a specific high impact business problem and build a proof of concept to demonstrate the value of agentic AI quickly, but at the same time also have a very long term vision for how this technology can really fundamentally wire your organizations, your people, people upskilling them, potentially even taking a dual approach of tactical execution. And then this is strategic vision. That's really the key to not just adopting agentic AI, but also truly competing and thriving with it.

Speaker A: Wonderful. Unmesh Maruti Sala uh, I think you've given our listeners some really great piece of advice. Unmesh take a measured risk. Maruti have that foundation philosophy and Salah, uh, start small but think big. And I really want to thank everyone for tuning in today for this beyond possible podcast Bharuti Unmesh Salah this has been fantastic. I look forward to having you all next time and I really appreciate your time today. Assuming one of these new AI agents doesn't take my job, I will be your host again. In the meantime, be sure to please check out Unmesh and Maruti's blog@treatance.com and have a great rest of your day. Thanks everyone.

Speaker C: Run. Thank you, thank you.

Speaker B: Thanks Alex. Thanks everyone.