The agent that reports the most is rarely the one shipping the most
Nate B. Jones — a practitioner who writes a daily AI strategy newsletter — published a video this week titled "Are Your AI Agents Doing Real Work Or Productivity Theater? Here Are 4 Principles I Use." The companion piece on his newsletter describes a failure mode that doesn't feel like failure. You asked an agent to watch for something, and it keeps telling you nothing has happened. Another thread finished its assignment and came back with six more things you could do. A third one, playing chief of staff, collected everything into a beautifully organized list — which you now have to read.
Every individual update seems reasonable. Together, they become a job. You delegated the work and came out the other side with more work.
The kicker: the agents aren't malfunctioning. They're doing exactly what they were told. The problem is that nobody decided whether the output was worth reading in the first place. As agents get cheaper to run and easier to spin up, the gap between looking productive and being productive gets harder to see, not easier — reports nobody reads, parallel drafts nobody compares, recurring jobs nobody remembers starting. The video's name for it is productivity theater, and the operator feels busy the whole time.
Why this bites a small business hardest
Every agent result creates a second job: a human has to read it, judge it, and decide whether it matters. In a big company, that second job is someone's actual role, with a title and a budget. In your business, it's you — after close, after the customers, after the work the report was supposedly about. Three standing agents adding ten minutes of reading a day is a part-time manager you never interviewed, paying you in summaries.
And the theater is convincing, because agent output looks polished by default. The plan is formatted. The prototype runs. The weekly research brief keeps arriving, still accurate, still relevant to a decision you already made months ago. None of it tells you whether the real work moved.
That's the audit nobody hands you: what your agents did, versus what they say they did.
The weekly truth-teller review (20 minutes)
Run this once a week, same slot. Four moves.
1. Pull the receipts, not the reports
For each standing agent task, write two lines: what it reported this week, and what actually shipped. Don't ask the agent for its version — ask the record. The invoice went out or it didn't. The page is live or it isn't. The reply got posted, the fix got deployed, the lead got called. The gap between those two lines is the entire finding.
2. Run the truth-teller pass with an AI outside the working chain
Hand those two-line records to a different AI — not the ones that produced the work — and ask one question per task: did this get far enough to matter? Jones's truth-teller prompt, described in his newsletter, is built for exactly this: it tells you what to finish and what to stop. The outsider matters. An agent reviewing its own output has every incentive to grade its homework generously; a separate one, given only the receipts, has none. It surfaces which stalled tasks deserve one more push and which are dead weight wearing a progress bar.
3. Give the summarizer an attention budget
If you run a manager-style agent that hands you a daily list, know what it's optimizing for: a complete, tidy list — not whether any item deserves you. Jones's principle here is that a chief-of-staff agent needs an attention budget, not just a task list: a standing instruction that your attention is the scarce resource, that decisions arrive with a recommendation and the evidence behind them, and that items get surfaced only when acting on them is worth the reading. Ten minutes of your Tuesday is not free. Make the list-builder price it.
4. Put a stop condition on everything that survives
Any recurring job that passes the audit gets an end date now, not later. "Run for three days and stop if I don't respond." A watch that speaks up when the answer moves, instead of daily to confirm it hasn't. Jones's example of the failure is a weekly market-research brief that keeps arriving for months after you already chose a different market — still accurate, still irrelevant. The fix isn't remembering to cancel it; the fix is building the stop in at creation, so continuation becomes a fresh decision instead of a default.
And one rule for your side of the table: an agent surfacing something unexpected is fine. Letting its curiosity set your agenda is not. The decision to chase a new lead belongs to you, not to the system that generated it.
Honest boundaries
- Weekly, not daily. A daily output audit becomes its own productivity theater. Twenty minutes once a week is the dose that survives contact with a real business.
- Stalled isn't the same as theater. Some weeks the real work legitimately couldn't finish. The audit distinguishes "slow" from "spinning" — killing a task that merely had a bad week is how you end up doing everything yourself again.
- The audit is only as honest as your outcome records. If you can't state what actually shipped this week, write the shipped list first, before the agent-reported list — otherwise the truth-teller pass turns into vibes about vibes.
- This is a pro-delegation discipline. The point of killing the dead jobs is that the live ones get your full attention. Audit so you can delegate more, not less.
Tonight's version
List the three standing agent jobs that have run the longest. For each, write the two lines: what it reported this week, what actually shipped. Kill the weakest one. Put a stop condition on the other two. Twenty minutes tonight — and next week's version takes ten, because you'll already know where the gap lives.
Want the templates? The Boring Automation Pack includes the delegation worksheets and the stop-condition templates — recurring-job end dates and escalation rules, ready to paste into your next agent setup.
Top comments (0)