Ken Thompson added pipes to Unix in 1973, in Version 3. A pipeline chains programs so the output of one becomes the input of the next.

A shell pipeline is fixed by the person who types it. An agent composes its own at run time, because the model reads the result of one tool call and decides which call comes next.

A single input can therefore take a different path on the next run, and the code around the model has to expect that and recover from it.

The engineering in an agent sits outside the model, in the harness that runs the loop and the handling for when it goes wrong.

What an agent is

The loop and the harness

Protocols

A single agent or several

Failure

How do the parts of an AI agent fit together?

The model chooses the next tool call, the loop carries the result back to it, and the harness runs the loop. The agent loop page covers the four phases, and the agent harness page covers the layer around them.

The loop's remember step has a topic of its own, and AI agent memory covers what gets written and read back.

A coding agent is the agent a developer uses daily, and the coding agents topic covers the files that configure one.

Each turn of the loop spends tokens and latency, which the LLM inference topic measures.