Agents Are Just LLMs Running Tools in a Loop

If you run an LLM in a loop and most LLMs are similar, what makes different agents have different capabilities? The answer lies in tools, workspaces, and specialization.

The Core Insight

Simon Willison's definition of an LLM agent has become the industry standard: "something that runs tools in a loop to achieve a goal."

At Decisional, we started seeing this when we moved from RAG to Agentic RAG. The performance felt mind blowing and allowing the Agent to do runtime reasoning made our system have sub 1% hallucination rates.

What Differentiates Agents

The answer can be understood by comparing agents to human employees:

• The Brain: The AI model—some general purpose, some specialized
• Tools: Capabilities like search, editing, code execution
• Workspace: The environment where the agent operates

The Evolution

In 2023, everyone wanted to build Jarvis. They would demo systems that worked 80% of the time which wasn't good enough. Models got better and trained to be great at calling tools in a loop—specifically ChatGPT o3 and Claude Sonnet 4.

Engineering Over Vibes

Agent building is moving towards more engineering than vibes and that's a great thing. The agents that work in production are the ones built with software engineering discipline, not just clever prompts.