AI Agent Weekly Report: What to Demand Every Friday
A good Friday report from your AI agent has five sections and leads with numbers the agent cannot fake. A weekly agent report is a fixed file template the agent fills in every week, showing what shipped, what stalled, the source-backed numbers, the decisions waiting on you, and next week's plan.
Most owners never define the report, so they get whatever the agent volunteers - usually a cheerful list of activity with no losses and no verifiable figures. The fix is to demand a set shape and the same evidence every week. This is the report the agent hands up; the matching ritual on your side is the weekly review you run as the owner.
What belongs in the Friday report?
The Friday report belongs in five sections, always in the same order, always present even when a section is thin. The five are: what got done, what did not work, the numbers, decisions waiting on you, and next week's plan. Keeping the order fixed is what lets you skim the same spots every Friday instead of hunting through prose.
Section one, what got done, lists shipped work with links - a deployed page URL, a commit hash, a published product - not a description of effort. If a line has no link or artifact to point at, it is not "done" yet and belongs in section two. Linkable output is the difference between a report you can check and a report you have to trust.
Section two, what did not work, names where the agent is stuck, what it tried, and what it is blocked on. A report that only lists wins is hiding the exact information you need to help. Treat an empty section two as a warning sign, not a triumph, because real weeks always contain friction.
Section three, the numbers, gives each figure with its source so you can re-pull it yourself. Section four, decisions waiting on you, is the short queue of things the agent cannot do alone - a spend, a price change, a publish - each with the option it recommends. Section five, next week's plan, states the top few priorities so you can redirect before the work starts, not after.
| Section | What it answers |
|---|---|
| 1. What got done | What shipped this week, with a link or artifact for each item? |
| 2. What did not work | Where is the agent stuck, and what has it already tried? |
| 3. The numbers | What moved, and what is the source for each figure? |
| 4. Decisions waiting on you | What needs your approval, and what does the agent recommend? |
| 5. Next week's plan | What are the top priorities before the next session starts? |
Because the report is the same five sections every week, you learn to read it in a minute or two. For the wider habit of skimming an agent's output without reading everything, see how to review AI agent work fast.
What numbers can't the agent fake?
The numbers the agent cannot fake are the externally verifiable ones - figures that live in a system outside the agent, where you can check them without asking the agent anything. Three carry almost all the weight: git commit history, live sales from the store's API, and hosting analytics pageviews. Each of these has an independent source you can open yourself.
Git commit history is timestamped and diffable: you can see exactly what changed, when, and by whom, and no summary can inflate a diff that is not there. Live sales pulled from the store's API are real money in a real account, not a claim about interest. Hosting analytics pageviews come from the platform's own counter, so traffic is measured by the host and not narrated by the agent.
Contrast those with activity numbers the agent can inflate at will: words written, tasks touched, files edited, hours "worked." Activity numbers feel like progress and measure motion, not outcome - an agent can produce a large word count and zero shipped value in the same week. When section three of the report leans on activity instead of external sources, that is the signal to ask for the source, every time.
The rule for section three is one line per number, and each line names where you can re-pull it: this many commits (git log), this much in sales (store API), this many pageviews (hosting analytics). A figure with no re-pullable source is an opinion wearing a number's clothes, and it does not belong in the report.
The last thing that makes the report trustworthy is its format. The report is a file template the agent fills in - it lives in the state file and the daily log, not in a one-off chat prompt. A prompt is forgotten the moment the session ends, but a template in the files is read and rewritten every session, so the report survives dead sessions and stays comparable week to week. Same five sections, same three sources, and what changes between two Fridays is the trend, not the tone.
FAQ
Should the report be a prompt or a file?
A file template, not a prompt. A prompt you type into chat is forgotten the moment that session ends, so next week's agent has no idea what shape you wanted. A template that lives in the state file and the daily log is read and updated every session, which means the report survives dead sessions, keeps the same five sections, and stays comparable from one Friday to the next. Put the shape in the files, not in a message.
How do I compare one week to the next?
Keep the same five sections and the same source-backed numbers every week, so the only thing that changes is the trend, not the tone. When section three always reports commits from git, sales from the store API, and pageviews from hosting analytics, this Friday's figures sit directly beside last Friday's and the direction is obvious at a glance. Comparability comes from a fixed template and fixed sources, not from how upbeat the writing sounds.
What about a week with no real progress?
The report must still show it. A Friday report with only wins and no stuck items is an incomplete report, and that absence is itself the red flag. A genuine week has friction, so section two should name what stalled and why; a report that hides it is worth less than one that admits a flat week honestly. No progress is a valid outcome to report - a hidden lack of progress is the actual problem.