The spec for the seat in the middle of your chart. One routine that gathers what happened, sorts it by what needs your decision, and hands you a short brief. The output is a specification you could hand to a builder, not a running routine.
Use it after you have drawn your org chart and before you stand up any specialist. Use it if you already have several agents and find yourself checking each of them separately, which is the symptom of a chart with no coordinator.
Do not use it to design agents that talk to each other. That is the thing this deliberately does not do.
One real week. What landed in your inbox, what you checked, what you decided, and what sat waiting because nobody noticed it had finished. Be specific about the waiting, because the handoff is where small operations lose the most time.
You are helping me specify ONE coordinating routine for my business.
It is the seat between me and everything else. Its job is not to let
agents talk to each other. Its job is to decide what reaches me.
Do not write code and do not recommend tools.
STEP 1. Ask me to walk through one real, recent week: what arrived,
what I checked, what I decided, and what finished and then sat there
waiting for someone to notice.
STEP 2. Sort everything I told you into exactly three buckets and
show me the sort:
A. Needs my decision
B. I only need to know it happened
C. Nobody needed this and I looked at it anyway
Be blunt about bucket C.
STEP 3. Write the spec, in plain language, under these headings:
GATHERS what it collects, and from where
SORTS the rule that separates A from B, stated so a
stranger could apply it the same way twice
ESCALATES what reaches me, in what form, and how often
IGNORES what it deliberately drops, named explicitly
STOPS the condition under which it should do nothing
and say nothing
STEP 4. Tell me the smallest possible version of this that would
still be useful. Assume I will build the small version first.
End with one sentence naming what this routine must NEVER decide
without me.
Run it once you have the chart. Twenty minutes. Build the smallest version it describes before you build anything else in your chart, because the specialists have nothing to plug into until this seat exists.
The IGNORES and STOPS sections matter more than they look. A coordinator with no stop condition becomes another thing generating notifications you learn to skim.
A sorting rule specific enough that you could apply it twice and get the same answer. A named list of things it drops. A stop condition. And a last line naming what it must never decide alone, which should make you slightly uncomfortable if it is honest.
Letting everything land in bucket A, which means you have specified a notifier rather than a coordinator. Skipping bucket C because it is unflattering. Leaving IGNORES empty. Specifying the full version instead of the small one, then never building it.
The value of this seat is the escalation, and the escalation ends at a human. That is the part with evidence behind it. My own system handed me four confident, wrong answers in a month, all traced to real files. No amount of agents conferring would have caught that. A person checking did.
Source: Build to Thrive, The AI Blueprint, week of August 17, 2026.