Tool

The dependency tower

The five conditions are not a list with ticks, they are a tower: each one rests on the one below it. Pull a block out with a click and watch what falls on top of it.

the data is written in one place onlythe process is written downthere is a handoff criterionthere are permissions and limitsthere is a way to measureclick a block to pull it out or put it backbase

The tower is whole. All five blocks in place: the process can hold an agent. Now pull the bottom one out and watch what happens to the other four.

The order is not decorative. It runs bottom to top in order of dependency, not of importance.

The model, in the open

The five blocks and the order they rest in come from the post this tool accompanies and from what we see at the start of an implementation. The tool deliberately gives no score out of five: four out of five is not 80% when the missing one is at the bottom.

How to read it

  • Pull the bottom block out first. It is the one least argued about in a meeting and the one that takes the most down with it when it is missing.
  • Notice that the blocks that fall stay on screen, tilted. They are still there, holding nothing up: which is exactly what happens to a process that has measurement and permissions but writes the same data from two places.
  • Put the block back and jump to the second one. The damage changes shape depending on which one is missing, and that is the argument against the list with ticks.

This object is the interactive half of an article: What has to work before putting an AI agent into a process

The playbook already exists. We multiply it.