Agent Anatomy

An agent’s anatomy is three parts: instructions, a model, and tools. Instructions are the text that tells the model who it is, what the job is, and what rules bind it. The model reads and decides. Tools are the functions it may ask to run. A coding assistant, a support bot, and a research pipeline all reduce to this triple; the loop runs the three against each other until the job is done.

The triple is a checklist for evaluating agent products. A framework, a platform, an “AI employee” is packaging around the same three parts, so ask where each part lives and who may change it: can you read the instructions, swap the model, see the tool list? A product that hides one of the three still contains it; hiding only removes your access.

It is also a debugging map. A misbehaving agent fails in one of three separable places: instructions that underspecify, a model too weak for the decision, or a tool whose description promises something its code does not do. Check the instructions first, because they are the cheapest of the three to change.