Peeling the Agentic Systems Onion: 5 Things to Consider

When I wrote my thesis on agentic systems twenty three years ago, I lived in a simpler world! No intelligence, no compute and no code to rely on to build scaffolding for my work. I built all my agents from scratch. Not only did these agents successfully accomplish their tasks, but in doing so, I came to fully appreciate both their capabilities and their limitations. That hands-on knowledge proved critical in deciding when and how to rely on them.

Today, it’s easy to get swept up by demos and off-the-shelf frameworks, but there’s real value in understanding the nuts and bolts behind the scenes - especially when it comes to agentic systems leveraging LLMs!

Here are five things to keep in mind before adopting the latest and greatest agentic “infrastructure.”

1. Keep It Simple

“Don’t get lost in abstraction.”

It’s tempting to jump on the newest framework that promises everything out of the box, but overly complex systems can obscure vital details. If you’re prototyping—or need to deeply understand how your agent is making decisions—start with a minimal design. Building from scratch, or close to it, helps you know exactly why your system behaves the way it does, making it easier to troubleshoot and iterate.

2. Verification in Regulated Environments

“Complexity comes at a cost, especially when validation is key.”

If you’re in a regulated industry—finance, healthcare, defense, etc.—you need transparency and solid evidence your system behaves as intended. Complex architectures can make validation a nightmare, making it hard to certify that every step meets stringent guidelines. Keep the architecture straightforward enough that you can demonstrate compliance and explain results.

3. Replicability & Auditability

“Follow the bread crumbs.”

A system that’s too opaque can become a black box. Good agentic architectures ensure you can repeat the same process and get the same results (replicability) and can trace how those results were reached (auditability). When building or choosing your agentic framework, make sure you can track decisions at every layer. This visibility is essential for troubleshooting and building trust in your system.

4. Dissect Every Code Example

“One size never fits all.”

Online examples can be a great starting point, but they often gloss over edge cases or are tailored to very specific needs. Before you lift code directly, break it down to understand each part. Confirm the assumptions it’s making, see if those align with your data and goals, and adapt accordingly. Just because something “does the job” in a demo doesn’t mean it’s the best solution for your unique situation.

5. No One Understands Your Use Case Better Than You

“Responsibility ultimately falls on you.”

Agentic infrastructures may simplify some complexities, but their designers don’t know your domain, constraints, or success criteria. You’re the one accountable for outcomes—so do the necessary due diligence. If you don’t fully understand how an agentic framework works, it can lull you into a false sense of security. Take the time to validate and customize accordingly.

Conclusion:

Agentic systems can be incredibly powerful, but a solid grasp of their mechanics and limitations is crucial. When I built my own agents from scratch, I learned first-hand why and when these systems can help or hinder a project. Start simple, keep it transparent, and never lose sight of your project’s specific needs. That’s how you peel the agentic onion—one layer at a time.

Best,

Sri Krishnamurthy, CFA, CAP

QuantUniversity

#AgenticSystems #AIArchitecture #TechInnovation #DataScience #MachineLearning #LLM