An agent can build fast, but it can't mark its own homework. That's my job.
with Elizabeth Cecilia. About 8 minutes. Tap any for its meaning.
Predict first
"Done. All good."
You asked an to add an information button to every tile in a Power BI report. It edited 104 files in two minutes and reports back: done, all good.
What do you do next?
The idea
The agent writes. Something else checks.
A is a small program that reads the agent's output and checks it against the : the written description of what should exist. It only reads. It writes nothing. It ends with one word, PASS or FAIL, and an that other tools can act on.
The point is the split. The agent is : ask twice, and you may get two slightly different builds. The validator is : give it the same files and it gives the same answer. You let the flexible thing do the work and the rigid thing do the checking.
In the project behind this series, the validator is one Python file. It runs 1,533 checks on every build, reads the model and report files plus the spec, and never writes a byte.
Sort it
Agent or validator?
Tap who each line describes.
Writes and edits the files
Gives the same answer every time for the same files
Might do the task slightly differently if you ask again
Answers with a line saying which file failed and why
Can explain its reasoning in plain English
Explore
The validator is one part of a harness
The validator doesn't work alone. It's one in a : everything around the agent that steers and checks it. Tap a station, or play the loop.
Tap a station
Five stations, one loop
Each change goes round once. The agent is only one stop.
Take the validator out of the loop. What catches a wrong file now?
Predict, then run
Break it on purpose
Here's a toy report: three tiles, each with an information button in a group, built from one small table. Its validator runs 29 checks. Pick one thing to break, predict how many FAIL lines you'll see, and the validator runs.
Pick something to break.
Check
What PASS means
The validator prints PASS. What do you know for sure?
Recap
Three things to keep
1An agent's "done" is a claim, not a check. The validator is the check.
2The agent is flexible and ; the validator is rigid and . Each does the job the other is bad at.
3PASS means "matches the spec, for what it checks". It's one , not the whole .
Everything is free on GitHub. Open the report next to the lessons and check each step for yourself.
The whole project
The Power BI project, the seed project, the specs and the tools: the whole kit. On the GitHub page, click the green Code button, then Download ZIP, and unzip it anywhere.
The Critical Raw Materials Intelligence report as a single .pbix file to open in Power BI Desktop. Open the latest release under Releases, then Assets, on the right of the GitHub page.