Read 420 meeting notes. Find the decisions nobody wrote down.
numbers a model wrote: 0 · step: cut
×
Text version
What the picture shows, for people and for AI. Example job, invented numbers. The pattern is from the guide's README, section "Agents: one craft each".
When a job is too big for one agent, it runs like a production line. One rule holds it together: the models read and judge, the code counts and certifies. A number in the report is never one a model wrote from memory.
In the picture, round characters are models and square ones are code. The example job: read 420 meeting notes and find the decisions nobody wrote down.
Cut (code). A small script is the foreman. It reads nothing: it cuts the 420 notes into 7 packets of 60.
Read (models). Seven worker agents read their packets in parallel and mark every note: a decision was made here, or none. Code checks each answer before it is accepted; 7 of 7 pass.
Sort (code). 332 notes had no decision and stop there. Only 88 go on.
Judge (model). One judge groups the 88 notes into decisions. It has one order: write no numbers.
Check (code). A gate checks the judge's answer: every note it cites exists, and every quote is really in its note. It passes. If it failed, nothing would leave the line.
Count (code). Code counts 31 decisions, finds 18 that are in no written list, and writes the report. Numbers written by a model: 0.
Why code does the counting: in a real case from the guide, a model judge reported 29 items and a script re-counted 17. The script was right.