Guides are easy. Sensors are the job.

The first thing every team does with coding agents is write instructions: an AGENTS.md at the root, a folder of rules, a few skills. It takes an afternoon, the next session is visibly better, and it feels like the problem is solved.

It is half solved. Birgitta Böckeler’s harness engineering article names the two halves. Guides steer an agent before it acts. Sensors observe what it did and help it correct. Her warning about guides on their own is exact: you get an agent that “encodes rules but never finds out whether they worked.”

In our experience, guides get the investment and sensors get whatever is left. A guide is a markdown file you can write in an afternoon. A sensor is a test, a check or a review agent that runs on every change and tells the agent what broke and where, and it takes far longer to build.

The sensors are where most of our work goes. The full post will cover why guides win by default, what a sensor looks like when it has to check hundreds of changes a week, and how a mistake an agent makes becomes a new check instead of another line in the rules.