Agentic AI Development: Agent SDKs & Multi-Agent Orchestration voor EU-conformiteit

[0:00] Welcome back to EtherLink AI Insights. I'm Alex, and today we're diving into something that's fundamentally reshaping how enterprises build AI systems. If you're tracking the AI space, you've probably noticed that chat bots feel like yesterday's news. In 2006, we're seeing a massive shift toward what's called a gentick AI, these autonomous digital co-workers that can handle complex, multi-step workflows. We're going to explore Agent SDKs, multi-agent orchestration, and how organizations in Europe [0:33] are building all of this while staying compliant with the EU AI Act. Sam, thanks for joining me. What's drawing your attention most about this transition? Thanks, Alex. What really stands out to me is the sheer magnitude of the shift. We're talking 73% of enterprises now prioritizing multi-agent orchestration up from just 28% in 2024. That's not incremental. That's a fundamental reconceptualization of what enterprise AI does. [1:03] But here's what keeps me up at night. 61% of EU companies still say compliance uncertainty is blocking their AI adoption. So we've got this massive technical momentum, but governance is lagging. That gap is exactly what today's conversation needs to address. That's a critical tension. So let's unpack what a gentick AI actually is, because I think a lot of people here agent and assume it's just a smarter chatbot. What makes a gentick systems fundamentally different [1:35] from what we were doing with prompt engineering and chatbots in 2024? It's night and day, honestly. A chatbot waits for you to ask a question and responds to that single query. An agentic system? It's autonomous. It decomposes a goal into sub-tasks, executes them without constant human intervention, handles multi-step reasoning with feedback loops, and crucially, it coordinates with other specialized agents. Think of it like the difference between having a smart assistant [2:05] who answers your questions versus having a team of specialists working toward an outcome while you're not even watching. Right. So autonomy is the key differentiator. And you mentioned McKinsey's research shows enterprise investment in agentic workflows is up 340% year over year. That's insane growth. But where are organizations actually deploying these systems? What are the real world use cases? The use cases are everywhere once you look. Invoice processing, customer support orchestration, [2:37] compliance review, knowledge synthesis, basically any workflow that's repetitive, multi-step, and requires coordination. There's a great example from Helsinki's financial services sector, an insurance company deployed a multi-agent system that reduced claims processing time by 62% while hitting 98.7% accuracy. The system had three specialized agents, one for document extraction, one for compliance verification, and one for risk assessment. [3:08] Each agent knew its lane, operated with clear boundaries, and left complete audit trails. That's exactly the kind of concrete example that matters. And notice Sam said, complete audit trails. That's huge for compliance. In Europe, the EUAI Act isn't just a suggestion. So when organizations are building these multi-agent systems, what's the technical foundation they need to avoid building on quicksand? That's where agent SDKs come in. Building agentic systems from first principles [3:40] is brutally complex. Your managing state, handling tool calls, controlling execution loops, managing timeouts and retries, and simultaneously maintaining compliance context. Modern agent SDKs, especially those built around frameworks like Etherlinks, AI, Lead, Architecture, give you structured patterns so you don't reinvent the wheel. OK, so agent SDKs are kind of like the scaffolding. What does that scaffolding actually provide? What's inside these kits that makes development faster? [4:12] Several core pieces. First, you get agent definition languages, basically a way to declare what an agent is, what it can do, and what its constraints are. Then there's a tool registry where you can safely register functions the agent can call with type checking and permission controls built in. You get state management for persistent memory and conversation history optimized for these long reasoning chains. And critically, you get an execution engine that handles the loop control, retry logic, timeout handling, [4:46] and interruption points. Without these primitives, you're building from scratch every time. So it's like the SDK handles all the operational complexity, which lets teams focus on the logic. But we keep coming back to this governance piece. You mentioned that 61% of EU companies cite compliance uncertainty. How do these frameworks actually address compliance, not just technically but operationally? This is where the conversation gets really important. Compliance in agentic systems isn't an afterthought. [5:17] It has to be baked into the architecture. You need observable systems where every decision, every tool call, and every agent interaction leaves an audit trail. You need fail safe boundaries, hard constraints that prevent an agent from doing something it shouldn't, plus escalation triggers that kick issues to humans when needed. So the compliance isn't just about documenting what happened. It's about building it so that violations are prevented in the first place. Exactly. And that's where production evaluation frameworks come in. [5:50] Before you deploy an agent system to handle real-world data and decisions, you need systematic ways to test whether it stays within its boundaries, handles edge cases correctly, and maintains the accuracy and fairness standards you've defined. The insurance example we mentioned earlier, they didn't just deploy and hope. They had structured evaluation against compliance benchmarks before going to production. That makes sense. So if I'm a tech leader at a mid-sized European organization and I'm thinking, OK, we need to move from chat [6:21] bots to agentic systems. What's the practical starting point? What's the roadmap? Start with a clear governance baseline. Before you write a line of code, understand what compliance requirements apply to your use case. That's your EU AI Act starting point. Then define your agents and their boundaries precisely. What's this agent allowed to do? What data can it access? What decisions can it make independently versus escalate? Then you select an SDK that aligns with your architecture needs [6:52] and build evaluation frameworks into your development process from day one, not after deployment. And the evaluation piece is that something organizations typically have in-house expertise for, or is this an area where external guidance helps? Honestly, it's mixed. Most organizations can define happy path testing. Does the agent do the right thing when everything works? But adversarial testing, fairness evaluation, and compliance scenario testing? [7:22] That's where a lot of teams benefit from frameworks and external input. You're not just testing whether your insurance agent approves valid claims. You're testing whether it's biased against certain demographics, whether it gracefully handles incomplete information, whether it escalates appropriately when it's uncertain. So we're really talking about a maturity shift, not just a technical shift. And I want to come back to the multi-agent orchestration piece because I think that's where things get really interesting. When you have multiple specialized agents [7:53] coordinating, the complexity compounds, right? It absolutely does. You're no longer testing a single agent in isolation. You're testing whether agent A correctly passes information to agent B, whether agent B's constraints are respected in that context, whether the system as a whole converges on the right answer. That's why orchestration frameworks are so important. They give you a way to define how agents interact, what information flows between them, and how decisions are made when agents disagree [8:24] or have conflicting goals. That's a really important point. And Helsinki's financial services example had three agents. How did they handle cases where those agents needed to coordinate or potentially disagreed? The system had explicit handoffs. The document extraction agent pulled data, passed structured outputs to the compliance verification agent, which then passed a compliance flag and any findings to the risk assessment agent. If any agent found a red flag, the entire workflow [8:54] escalated to a human reviewer. That clear, linear orchestration made debugging and auditing straightforward. In a more complex scenario, you might have agents that can request information from each other or even suggest alternative paths. But you'd want that negotiation logic to be explicit and auditable. So the lesson is that orchestration isn't just about throwing agents at a problem and hoping they coordinate. You have to design those interactions deliberately. [9:25] Right. And here's the bigger picture. If you want 98.7% accuracy with clear compliance trails like that insurance system achieved, orchestration has to be intentional. You're not just scaling chatbot logic. You're architecting a system where multiple specialized intelligences work within guardrails, with humans maintaining control and visibility throughout. Perfect. So as we wrap up, let me ask you this. If you're advising organizations right now about their 2026 AI roadmap, what's [9:57] the single most important thing they should be thinking about? Don't treat compliance as separate from architecture. The organizations succeeding with agentic AI are those that build governance into the design from the beginning. That means choosing SDKs and frameworks that support compliance, defining agent boundaries explicitly, and treating evaluation as a first class engineering concern. If you wait until you've built everything to worry about governance, you're going to have to rebuild. [10:28] That's great advice. For listeners who want to dig deeper into agent SDKs, production evaluation frameworks, and how Helsinki-based teams are implementing this, the full article is on etherlink.ai. You'll find concrete frameworks, governance checklists, and a lot more technical detail about how to actually build these systems. Sam, thanks for the deep dive today. Always a pleasure, Alex. This is a critical moment for European organizations. [10:58] The shift to agentic AI is real. The compliance requirements are real. And having the right frameworks makes all the difference. Thanks for listening to etherlink.ai insights. We'll be back next week with more on AI governance and development. Until then, stay curious.