Decoded #9: Agent Graphs

A recent post on X set off a wave of posts stating loops are dead and graphs are the future, blah, blah, blah. Here is what the buzzword actually means.

The old picture was one AI agent running a loop. You give it a goal and it plans, acts, checks its work, and tries again. That holds up until the task gets big.

An agent graph is a different shape. Instead of one agent doing everything, you wire several specialized agents together. Each 'graph node' is an agent with one narrow-focused job. Each job completion is a hand-off with a defined contract. Jobs flow through a structure you defined.

Think about how a clinic runs. You staff roles (intake, workup, prescribing, billing, etc.). You define hand-offs. The structure is what makes the work safe.

Here is the part worth holding onto.

A single looping agent lets you skip mapping your workflow. A graph makes you do that homework up front. Every step gets a name. Every hand-off gets a contract. When something fails, you know where. We think that visibility matters most when the answer touches a patient.

Useful frame: a loop asks you to trust the agent. A graph makes the agent show its work.