Managing an AI Agent: Delegation, Review and Reporting
Managing an AI agent takes about 10 minutes a day once three things exist. You need tasks written so they cannot be misunderstood, a review habit that reads four files instead of everything, and a weekly report the agent writes about itself. Get those three in place and the day-to-day cost of running an autonomous worker drops to a short check-in, plus a 30-minute weekly review that catches drift the daily glance misses. This page is the reading path to all three, in the order that builds the habit.
The common failure is treating an agent like a chat window: dropping vague requests in, then reading everything it produced to figure out what happened. That does not scale and it is exhausting. Management by files scales, because the agent maintains its own state, log, and inbox, and your job becomes reading a handful of specific files rather than reconstructing the whole session by hand.
How much of my time does an AI employee really need?
About 90 minutes a week in steady state: roughly 10 minutes a day of check-in plus a 30-minute weekly review, after a one-time setup of about 15 minutes. The daily check-in reads the state file and the human-inbox and answers anything the agent parked for you; the weekly review reads the logs for the week and looks for drift, stalled work, and metrics that are not moving. The numbers here are our own operating rhythm, not a projection. The detail lives in how many hours a week managing an AI employee takes.
Start here: the reading path
Read these roughly in order. The first few set up delegation and unambiguous tasks; the middle covers the review and reporting habits; the last few handle overnight work, handoffs, and measuring outcomes.
- Your first autonomous AI agent session: the 15-minute setup - office, mission, contract, then "read the contract and begin." The exact starting sequence.
- How to delegate to an AI agent: the 5-part task that comes back done - the delegation grammar: a verb with an end, scope, a success test, boundaries, and a landing place.
- How to write a task an AI agent can't misunderstand - the five parts of a complete task, with three before-and-after examples of vague versus complete.
- How to review AI agent work in 10 minutes - read four files, not everything: state, latest log, human-inbox, and git history, plus one red-flag rule.
- AI agent weekly report: what to demand every Friday - a five-section report the agent writes about itself, led by numbers it cannot fake.
- The 30-minute weekly review that keeps an AI agent honest - the ritual that catches drift, which daily check-ins never will.
- How many hours a week does managing an AI employee take? - the honest time budget, broken down by daily, weekly, and setup cost.
- How to hand off work between two AI agents - shared files plus a start protocol and an end protocol, so the vendor and the session boundary stop mattering.
- Can an AI agent work while I sleep? An honest answer - the two-minute evening setup, hourly checkpoints, and what a 3 a.m. crash actually costs.
- How to measure an AI agent: activity vs outcome metrics - the one-number rule and why an early zero is data but not yet a verdict.
- How to stop an AI agent going off track (scope creep) - separate ideas from execution: one task to completion, and new ideas parked in the backlog instead of built now.
- When your AI agent says "done" but nothing happened - why agents report intention rather than result, and the three falsifiable checks that catch it.
- What breaks when an AI agent runs unattended - four real failure modes and the rule each one produced, with git as the single best insurance.
- Cheap model or expensive model? The routing rule that cuts bills - route each task by the judgment it needs, and set an explicit model on every subagent.
- When should an AI agent delegate to subagents? - the jobs worth a subagent, and the rules that stop parallel agents from colliding.
FAQ
Can an AI agent work while I sleep?
Yes, for well-scoped work, and the safety comes from structure rather than supervision. You set it up in about two minutes with a clear task and a definition of done, and the agent checkpoints its state to files every hour or so, so if the session dies at 3 a.m. a fresh one resumes from the last checkpoint on "continue." What it must not do overnight is anything on the never-alone list - spending money, publishing, messaging people - which stays gated behind the inbox until you are awake to approve it.
What happens if I skip a week of reviews?
Not much, if the guardrails and files are in place, because the agent keeps its own state and parks anything risky in the inbox rather than acting on it. You come back to a readable state file, a week of logs, and a queue of decisions waiting for you - not a mess to untangle. The cost of a skipped week is slower course-correction, not damage; the structure is designed so that the worst case of your absence is stalled progress, never a runaway action.
How do I know the agent is doing a good job?
Judge outcomes, not activity, and read the numbers it cannot inflate. Git history, store-API sales, and hosting pageviews are facts; word counts and task counts are not. Pick the one number that actually matters for the current goal, watch it over a fair trial window defined in advance, and treat a log that shows honest failures as a good sign rather than a bad one. An operation that only ever reports wins is hiding something.