Agentic Mesh: Building Highly Reliable Agents

LLMs are getting bigger but not necessarily better. Specialized models, backed by deterministic orchestration and an agent-based architecture offer a smarter, more accurate, and much more reliable path forward.

Source: Eric Broda

Introduction

As LLMs capabilities increase, software and user expectations expand to take full advantage.

But as we race forward there is a constant balance — revery new Large Language Models (LLMs) brings reliability (inaccuracies and what we call hallucinations) increases, but we ask for more, and sure enough reliability drops. Rinse and repeat.

But is there a better approach?

My belief is that today we simply ask far too much of LLMs. No sooner do we get a new, better, LLM and we ask them to do more, and once again we are confronted with errors and inaccuracies. So, when we build agents on top of LLMs, we end up inheriting the challenges of this LLMs.

Here is the crux of the problem: LLMs are probabilistic, but probabilities multiply and grow — sometimes catastrophically. So, when asked to do small things LLMs are extremely reliable and accurate. But when we ask them to do big things, we cascade small errors into exponentially growing errors. In a way, small is better.