Agent Wiring

Agent wiring is the choice between two ways to build an agent, and the cut is who owns the loop. In the first, you write a definition: instructions, a tool list, maybe a schedule, in a file a host reads and runs. The host owns the loop, the retries, the context handling; you own what the agent knows and may do. In the second, you write code against a framework: the loop is yours, the framework hands you the model calls and the primitives, and every behavior is yours to implement.

A real definition is small: one markdown file, a frontmatter block with the description that triggers it and the tools it may use, and the instructions as the prose below. A definition inherits the host’s loop, context handling, retries, and recovery, so when the host ships a better loop your agent improves without you touching it. The price is the host’s opinions: you shape behavior through instructions and tools, and where the host draws a boundary, your agent’s behavior ends.

The code route buys control. You can build a loop no host offers, test it like the software it is, and know what happens at each step. The price is that the loop is now your code: every improvement the hosts ship, you re-implement yourself, and every subtle loop-design bug, the kind hosts have spent years finding, is yours to find again.

Wire as a definition first and switch to code only where the host’s boundary blocks the job, because most jobs never reach it.