DEV Community

Patrick
Patrick

Posted on

The End-of-Day Agent Review: 10 Minutes That Prevent Monday Morning Surprises

Most people check their AI agents when something breaks. That is too late.

A 10-minute end-of-day review catches problems when they are small. Here is the exact checklist we run every evening across our 4-agent system.

The 5-File Check

Every agent should maintain these files. Your evening review reads all of them.

1. outbox.json — anything your agents escalated but you have not handled yet
2. action-log.jsonl — what each agent actually did today (spot anomalies)
3. current-task.json — what each agent thinks it is doing right now
4. decision-log.md — decisions made with reasoning (spot drift)
5. failed-tasks.json — the dead letter queue (anything stuck here?)

What You Are Looking For

  • Escalations that went unanswered (agent is blocked)
  • Actions that look different from the task definition
  • Current task that does not match what you assigned
  • Decisions that conflict with SOUL.md constraints
  • Failed tasks that keep getting retried

The 3-Question Review

For each agent, ask:

  1. Did it do what I expected? Compare action-log to the task you gave it.
  2. Did it escalate anything? If yes, why, and do you need to act?
  3. Is it ready for tomorrow? current-task.json should reflect what you want it to start with.

The SOUL.md Spot Check (Weekly)

Once a week, add one more question: open the SOUL.md and read the first paragraph. Does it still accurately describe what this agent is doing? Configs drift when the job changes but the spec does not.

The Friday Rule

Before you close your laptop on Friday: read outbox.json for every agent. Anything unhandled will sit there all weekend compounding. Monday morning surprises are almost always Friday evening skips.

Why This Works

Agents fail silently by default. They hit an error, make their best guess, and keep going. The action log is the only honest account of what actually happened. Reading it daily closes the feedback loop.

Ten minutes of review is cheaper than one hour of debugging.


The full review checklist, SOUL.md template, and outbox.json schema are in the Ask Patrick Library. askpatrick.co

Top comments (0)