[0:00] Welcome back to EtherLink AI Insights. I'm Alex and today we're tackling something that's reshaping enterprise technology, enterprise-agent AI, multi-agent orchestration, and what it actually takes to get these systems production ready. Sam, thanks for joining me. This feels like the conversation everyone in tech should be having right now. Absolutely. The stats are striking. McKinsey reported that 72% of enterprises are now evaluating or deploying agentic AI systems. But here's the catch. [0:32] 68% are hitting real deployment challenges, so we're not talking theory anymore. This is happening and it's messy. Messy is the right word. Let's start with the basics, though. What exactly is agentic AI? And why is it fundamentally different from the chatbots and single model systems enterprises have been deploying? Great question. Traditional enterprise AI, think customer service chatbots, relies on a single LLM to handle all the logic, memory, and tool calls. One model does everything. [1:05] That approach hits hard limits pretty fast. You've got context constraints, you can't really specialize the system, and when something breaks, it cascades everywhere. So it's a bottleneck problem. One agent trying to be good at everything ends up being mediocre at most things. Exactly. Multi-agent systems solve this by decomposing workflows. You have a planner agent that breaks tasks into sub-tasks. Specialist agents execute those sub-tasks in parallel. A coordinator makes sure dependencies are met. It's basically how human [1:39] teams actually work and the data shows it performs better. That makes intuitive sense. You wouldn't have one person doing sales strategy, deal execution, and customer success. You'd have a team. Why did it take us so long to do this with AI? Partly because the underlying models weren't sophisticated enough to coordinate reliably. And partly because most enterprises defaulted to existing frameworks and vendor solutions built around single agent paradigms. But as LLMs got better at planning and reasoning, [2:11] the architecture started clicking. Okay. So once you commit to this multi-agent approach, what are the actual architectural patterns you're working with? Are there best practices emerging? Gartner identified three dominant topologies. Hyerarchical where a planner delegates to workers, very common in supply chain and HR. Swarm topology where peer agents collaborate without a central controller, great for discovery and brainstorming. And pipeline where one agent's output feeds the next, standard for content generation and data processing. [2:46] And in the real world, do enterprises stick to one pattern or do they mix and match? Most production systems blend them. I'll give you a concrete example. An enterprise claims processing workflow might use hierarchical planning for case routing, pipeline logic for document extraction and verification, and swarm agents for fraud detection. You're picking the right tool for each part of the problem. That's smart. Now when you're actually building these systems, you've got a classic make versus buy decision. How are enterprises approaching the SDK choice? [3:21] There's Langchain, Claude, AutoGen, Cloud Platforms. The landscape is pretty crowded. According to GitHub's 2024 AI report, Langchain and Thropics, Claude SDK, and AutoGen, account for about 65% of multi-agent projects starts in Europe. AWS Bedrock agents and Azure AI agent service are growing for organizations already invested in those clouds. Each has real trade-offs. Walk us through those trade-offs. What's the Langchain story? Langchain is flexible with [3:54] a huge community, but it has a steep learning curve and no built-in evaluation framework, which as we'll get into is critical. And Thropics SDK has native tool use support and excellent documentation, but you're locked into their ecosystem. AutoGen supports multiple models and handles conversation management well, but it's less battle-hardened in production environments. Claude Platforms offer integrated logging and governance, but less flexibility and higher costs. So there's no obvious winner. The choice depends on your constraints. [4:28] Exactly. And here's what we've learned. Framework choice matters less than architectural discipline. The winning pattern is abstracting your agent logic from the SDK. It lets you swap frameworks if requirements change, and it ensures portability if you move cloud providers down the line. That decoupling is worth the upfront effort. That's smart architecture thinking, but there's also a custom build path. When would an enterprise decide to build agents from scratch instead of relying on a framework? There are specific scenarios. First, if domain-specific [5:04] statement is critical, think financial trading or clinical workflows where you need very precise control over how state evolves. Second, if you need multi-step reasoning with memory that spans days or weeks. Third, if tool calling requires complex permission or validation logic. And fourth, if you're serving more than 1,000 concurrent users and need fine-grained cost control. Those are pretty specific gates. It sounds like custom builds are the exception, not the rule. [5:34] They are. Custom agents typically increase time to value by 4 to 6 weeks, which is significant, but at scale they reduce operating costs by 30 to 40%. So for most enterprises, start with a framework and only go custom if the math compels it. Let's talk about something that came up earlier. Evaluation. You mentioned Langchain has no built-in eVal framework. Why is evaluation such a big deal in a gentic AI? Because multi-agent systems introduce new failure modes that single-agent systems don't have, [6:07] an agent might make a logical error in routing, a coordinator might misunderstand dependencies. Agents might contradict each other. Traditional metrics like blue scores or semantics similarity don't catch these problems. You need frameworks that evaluate planning accuracy, task completion, failure isolation, and team coordination. So you can't just run AB tests and call it done. Not nearly enough. You need structured evaluation at multiple levels. Do individual agents [6:40] achieve their narrowly defined goals? Do multi-agent workflows complete end-to-end tasks? Do they do it faster and cheaper than the previous system? Do they degrade gracefully when one agent fails? These require different evaluation methodologies. That's a lot of rigor. And then there's the regulatory layer. The EU AI Act is a big elephant in the room here. How does that impact enterprise-agentic AI deployment? It's substantial. The EU AI Act classifies AI [7:12] systems by risk level. Many enterprise-agentic workflows, especially in HR lending and supply chain, land in the high-risk category. That means mandatory impact assessments, extensive documentation, human oversight requirements, and regular auditing. These aren't bolt-on concerns. They need to be embedded in your development process from day one. So compliance isn't something you handle at the end. It shapes your architecture. Absolutely. You need governance frameworks that [7:43] log every agent decision, trace reasoning chains, document why certain actions were taken, and prove that humans are meaningfully involved in high-stakes decisions. That requires careful design of your multi-agent system, not just adding reporting on top afterward. Given all of that complexity, orchestration, evaluation, compliance, what's the practical playbook for an enterprise actually moving from pilots to production? Start with clarity on your architecture. Pick your topology, hierarchical pipeline or swarm based on your workflow, not the framework. [8:18] Second, choose your SDK thoughtfully, but abstract your logic from it. Third, build evaluation into your process early. Don't wait until you're ready to deploy. Fourth, map your regulatory obligations, especially if you're in Europe, and design compliance into the system. And fifth? Fifth, iterate in stages. Start with a low-risk proof of concept, prove the evaluation framework works, then gradually increase complexity and stakes. [8:50] Production readiness isn't a single moment. It's a maturity progression. That's solid guidance. Before we wrap one final thought, we're in a moment where agentech AI is moving from hype to deployment reality. What's the biggest misconception you're seeing in the market right now? That agentech AI is primarily about automation efficiency. Yes, you get efficiency gains, but the real unlock is capability. Multi-agent systems can do things [9:20] single-agent systems fundamentally can't. They can handle workflows that require specialization, parallel execution, and adaptive re-planning. That's a different problem category entirely, and understanding that shift changes how you invest. So it's not just faster. It's capable of things you couldn't do before. Exactly. And that's why we're seeing 72% of enterprises exploring this space. They're not just optimizing. They're expanding what's possible. Sam, thanks for walking through this with me. There's a lot of depth here, [9:52] and our listeners are going to want to dig deeper. If you want the full technical breakdown, evaluation frameworks, governance considerations, and a detailed SDK comparison, head over to etherlink.ai and find the complete article. That's your roadmap for moving agentech AI into production. Thanks for listening to etherlink.ai insights.