Set-up

Everything around the model that decides how an agent behaves: its instructions, its memory, its tools, and its checks.

The model is the same machine for everyone. What makes one agent useful and another maddening is the set-up: what it reads at the start, what it kept from before, what it can actually reach, and what stops it. Change the set-up and the behavior changes; leave the set-up and no amount of persuading will hold.

The idea is load-bearing because of where it sends the blame. A blocked agent is a set-up problem. A missed instruction is a set-up problem. So is an agent that keeps volunteering what you never wanted, or that cannot tell you the time, or that saved something you never asked it to save. Looking there first is the same blameless habit the canon asks of any system: cause sought in the arrangement, never in the person, and the standard improvement rather than perfection.

A Player+ keeps the set-up legible. Plain text you own, in a folder or a notebook: the charter, the memory, the logs, the venue cards. Ask your agent where each part lives and look at it. Know which layers you can edit (yours), which you can steer and not edit (trained habit), and which you can neither see nor argue with (the maker's instructions above yours, and the guard around the action). When something must never happen, put it in the enforced layer rather than the read one.

Legibility is the point. Everything an agent does should be auditable in a text editor: no opaque state, nothing about you held somewhere you cannot open.

Also called: setup Stands on: Agent · Language model · Charter · Memory (agent) Opens onto: Agent engineering · Safety check · Zero Blame Environment · Hosted service · Key In play: beyond Sources: Player+ Modules, Agent Engineering, How an AI Agent Works, The Agent Improvement Process · lionsberg.wiki: Zero Blame Environment, Continuous Improvement (unmarked). Open: none found.