Thirty of our agents stopped being agents on 3 November.
They were the narrow helpers under our eight lead agents, the ones each lead used to hand work to. Now each lead does that work itself, from a written playbook inside its own definition.
The identifiers stayed the same, so routing and the capability lists still name every one of them, and the original files went onto an archive branch. The agent definition files followed the same day, with the thirty former sub-agent files removed. The check that validates the setup changed too: it now expects two reviewer agents and thirty playbooks, and confirms each playbook has an owner, a file and a heading.
The two reviewers are the part worth copying. Our voice reviewer and our measurement reviewer were not folded into anyone. They stay separate agents, so when one of them checks a piece of copy or an analysis, the check comes from something that did not write it. When a reviewer shares a definition, a context and a set of instructions with the thing it reviews, it tends to agree with it. That is our judgement, not a measurement, and it is the reason we kept them apart.
It is a leaner team now, and the change is a week old, so we cannot yet say what it saves or what it costs.
For the wider setup, see how the AI specialist bench works and what our agents are not allowed to decide alone. The Team page lists the playbooks and the two reviewers.
If you run a team of agents, or a team of people, ask which layers exist to do work and which exist to check it. Fold the first kind when a written method will do the job. Keep the second kind apart.
Where do responsibilities blur in your own marketing setup? The Hexagram Diagnostic is a free 8-minute self-assessment that shows which of your six pillars to look at first.
