AI Build Lab / Learn · Agent Workforce · Week 4 · Module 05 · Details

More detail: Turn on Evy's daily check

Optional reading for this lesson. Everything you must do or decide is in the lesson.

Week 4 · Module 05 · Lesson 1 · Details

This page goes with the Week 4 lesson Turn on Evy's daily check (course/workforce/week4/lessons/anw-w4-l08.md). You don't need it to finish the lesson. Read it when you want the longer version.

When it runs

Once a day, before your first brief, while your ruling says it's on. If you have not set a time for it in the lesson on scheduling, nothing else starts it: a day you don't ask for a brief is a day without a check, and the next check covers the work since the last one. A folder under runs/ named with the date and evy-daily shows it ran that day.

The catch-up check

Your Chief keeps a list of every job that's meant to run on a schedule, with when it last ran, what it does (only reads, saves files on your computer, or sends and posts) and whether it's all right to run late. Before your first brief, your Chief checks the date and time and sorts anything missed:

Kind of jobWhat happens
Only the latest run mattersIt runs once now
Every run countsEach missed day runs, in order
Too late to be useful nowIt's skipped, and the brief says so
Sends or posts anythingIt never runs late on its own; it waits in Needs you for your word

Each run is tied to the time it was meant for, so nothing runs twice.

What Evy reads

  • Your teams' saved work under runs/.
  • A copy of your Langfuse records, saved on your computer by a fixed script. The copy also keeps records past the 30 days the free Langfuse plan keeps them. On Windows the copy isn't made yet: Evy grades from your saved work and says which steps it couldn't check against the records.
  • A record of your chats with your Chief, in Claude Code and in Codex, made by a fixed script that keeps no one's words. It keeps only what happened: who acted, what was handed to which agent, what was sent, and review verdicts. It hides secrets, email addresses and phone numbers. It marks what it can spot by itself: a correction, a hold, a clean send, a repeated question, an interruption, a long gap, something left hanging. Evy reads those marks as leads, never as rules.
  • Your rulings, filed by the Charter Steward. When you say "That's a rule for the <team name> team: ..." and the Steward files it, Evy adds it to its notebook for that team on the next check, in your words as the ruling quotes them.

What the script keeps stays on your computer and is never put in a save point.

What Evy may do without asking

  • Add examples to its notebook for a team.
  • Add an item to Kansa's watch list for a team. Each one appears in your brief with a one-word undo.

It suggests rule changes; you decide them.

"How your teams did yesterday"

One short part per team:

  • what ran
  • what passed, and what failed and why
  • what was added to Kansa's watch list, with its undo
  • what you corrected that's now in Evy's notebook
  • one suggestion

The weekly review

On Fridays, about fifteen minutes to read: what ran; how much went out untouched, lightly edited, rewritten or thrown out; which golden cases failed; how often you and Evy agree, and which way it's moving; examples waiting for review; and one change to make next.

When you grade a job yourself

Evy grades every job each day, including the few the counting script picks for you to grade. Your brief puts each under Needs you as "grade this one: <job>". At first it picks every job, then one in two, then one in five, as you and Evy keep agreeing. Until you save your grade and your Chief seals it, Evy's verdict says "held until you grade it" on your dashboard, in your brief, in the daily files and in Langfuse. A job whose Evy answer you already saw cannot be picked. Don't open Evy's folder for a job you're about to grade.

Ask The Professor to save your grade for only the rules you looked at, then ask your Chief to seal it. Evy has already graded the job without seeing your grade. The seal compares the grades right then, shows Evy's verdict beside your reasons, counts how often you agree and notes that Evy graded first while its answer was hidden from you.

When a fix comes to you ready

Each problem Evy finds becomes one problem card, handed up to your Chief of Staff. Your Chief picks the kind of problem from a fixed list and writes down why, with Evy's quote; a script sends it to whoever's job the fix is and does the counting. The same problem seen twice is one fix, not two. At most two fixes a day get ready for one team, and three across your teams, with serious rules first; the rest wait for the next day. A problem you skipped or undid stays quiet until a new live job fails the same way, and one that two fixes didn't solve comes to you as your call (it may need the Architect (Archie)).

Whoever gets a fix ready writes only in that fix's own folder, and a script checks that nothing on your team or in your library changed. Your yes is to exactly what was made ready: if something it touches changed since, it says it's out of date and changes nothing. Fixes for one team are made one at a time. After your yes, the golden cases it affects rerun, Evy grades them from an order a script writes, and you see the before and after. Then you say keep or undo, and the change record gets its write-up either way. If you said "yes F4, keep it if everything passes", you only answer once: it's kept when every rerun case passes, and you're asked again if anything fails. To always be asked twice, say "Turn off the keep-if-everything-passes shortcut. Have the Charter Steward record my ruling." A fix nobody answered for 7 days expires.

On your dashboard, the Today tab has a "Fixes ready" card with the same fixes. Its buttons only copy the words to say to your Chief, like "Yes F4". A fix that touches a job you're still grading doesn't show there at all.

Why it stays on your computer

Evy reads your chats with your Chief, and those never leave your computer. Cloud options are on your "what's next" page in "Move your organization from class to work".

Staff only: internal notes. Not student-facing.
  • 12 October (L3, Tyler ruled): Evy grades every job; the owner's sample is hidden from the owner until saved and sealed (workforce/runtime/evy_hold.py); the seal runs the compare.