Harness, Loop, & Graph Part 3: A Simple Explanation of How AI Agents Are Built

I keep hearing people say that an infinite token budget solves all problems. Nowhere is that more true than with graphs. If tokens were free or if you are lucky enough to have hyperscaler compute levels, you can brute-force your way through many architectural sins.

Unfortunately, most of us lack the $100 billion hardware budget required for us to be lazy. Even Microsoft is telling its engineers to tighten their belts and optimize everything. We need graphs managing our agents, and building graphs without an infinite token budget requires us to solve the context and verifier problems.

Graphs aren’t new, but they deliver new capabilities, more control, and improved reliability. They also introduce a new problem: the learning problem. Loops take on a new role to support early iterative learning. Building, improving, scaling, and maintaining a graph manually is more expensive than infinite tokens.

Again, we need a bridge.

Graph engineering wires multiple specialized agents or steps into a graph data model. Nodes are the agents or steps. Edges are the routing between them, including branches, splits, merges, and loops. Shared state flows along the edges. Surprise, part three of your agentic harness is a data model and the tipping point into information models. Those two are different, and I will explain the differences in detail in just a bit.

With graphs, instead of one agent cycling through an entire bug fix, you build five steps. One reads the bug report and identifies which files are involved. Three, each examining one file, run in parallel. One collects their findings and writes the fix. Each step is its own model call with its own instructions, and the wiring between them is ‘coded’, so the model doesn’t have to make a decision with incomplete context.

A single loop is the simplest possible graph. It’s one node with an edge pointing back to itself. Workflows are more complex graphs.

You probably noticed that I put coded in quotes. It’s the wrong approach here, but it’s the default approach for software engineering. Logic lives in code, not your data model. Many are rediscovering the ‘code is logic and logic is symbology’ approach to agents.

However, information models store logic more efficiently, and graphs are the bridge from data models to information models. That means software is no longer our primary logic store. This is an entry point into one flavor of neuro-symbolic AI. I explained the concept further in a previous article, but don’t go there until you read the rest of this article.

You need to see a core concept upfront to understand how graphs for agents create the bridge we need.