Source: raw/Are_Your_AI_Agents_Doing_Real_Work_Or_Productivity_Theater_Here_Are_4_Principles_I_Use..md
Creator: Nate B Jones, AI News & Strategy Daily · URL: https://www.youtube.com/watch?v=NIRFVjttNvA · Duration: not recorded · Platform: YouTube (auto-captions, fetched 2026-11-01)
Jones argues that a lot of agent output is “productivity theater”: it “looks like work, it smells like work, it burns tokens like work, and it ends up on a board slide that nobody cares about.” He derived four principles from how he manages his own agents in Claude and Codex. All four come down to the human stating intent and limits, because the agents will keep producing as long as they are asked. A companion “truth-teller” prompt and a chief-of-staff write-up are on his Substack; neither was part of this source.
Key Takeaways
- 1. Count the human work after the agent finishes. Judging output by “how hard it would have been without AI” says nothing about whether you should spend time on it. A report has to be read, believed and acted on; sent to a team, that cost multiplies; made weekly, it becomes “a weekly claim on human attention.”
- 2. Run more experiments than you commit to. Use parallel agents to explore different explanations of one problem, not to build several solutions before you understand it. “Five agents producing five versions of the same answer” leaves you a comparison job that teaches little. When the missing fact is outside the agents’ reach (will customers pay?), go ask people.
- 3. Decide when an assignment ends before making it recurring. A recurring agent “can outlive any reason you created it.” His own example: a weekly market update that kept arriving two months after he chose a different market. Give stop conditions up front, for example “if I don’t respond to this email in 3 days, just stop sending it.” That is “not actually an agent problem. It’s a you problem.”
- 4. Give the agent that manages your work a budget for your attention. Not a number of minutes: a chief-of-staff agent should know your commitments and what you care about, screen what other agents raise, and “prepare the choices”, not pass everything through. “The agents have just about infinite attention… You cannot.”
- Don’t let agents steer your curiosity. “Make sure that you are the one that wants to go farther… not the agent.”
- Three audit questions for the agents you run now: which help with something you care about; which are experiments you are still learning from; which are “only running cuz you forgot to tell them to stop.” If you can’t answer, you may be doing productivity theater.
Examples he uses
- Dan Ninarol, Modern Classical Chefs (South Jersey). Used Claude to prepare proposals that he says brought in $20,000 in a month and a half. The follow-on work (cooking for paying customers) was work he wanted. The spelling of the name is from auto-captions.
- Tony Severino, painting contractor near Chicago (in business since 1984). Had Claude read PDF blueprints floor by floor for bids, caught it using a floor-area multiplier instead of real wall measurements, and now asks it to show its work. A polished proposal that skipped that check “would have been worse than useless.”
- Customers who try once and leave (hypothetical). Good: separate agents read cancellation reasons, help requests, and what returning buyers buy differently. Theater: one agent designs a loyalty programme, another a pricing page, another a win-back email, before anyone knows why people leave.
Try It
- List every agent and recurring job you run and sort each into Jones’s three questions: cares-about, still-learning, forgot-to-stop. Stop the third group today.
- Add an end condition to every recurring task when you create it: a date, a number of runs, or “stop if I don’t respond in 3 days.”
- Before accepting an agent deliverable, name who reads it and what happens next. If nobody, don’t commission it.
- For high-stakes numbers, ask the agent to show its work, or assign a second agent to “find the assumption that would change the answer.”
- Write a brief for your chief-of-staff agent listing what may interrupt you, which commitments must not slip, and what kinds of new evidence would change a decision.
Open Questions
- The truth-teller prompt and the chief-of-staff write-up were not reviewed.
- The two business examples are relayed by Jones; their original sources are not given in the transcript.
Related
- Your Agents Are Roommates (Nate B Jones) — the machine-resources side of the same “give every agent a stop sign” advice
- Recurring-Task Automation (Nate B Jones) — a recurring job worth automating, done with root-cause analysis
- Agent Lying Is an RLVR Artifact (Nate B Jones) — why agent reports need checking before you act on them
- Ryan Carson’s Clawd Chief — a working executive-assistant agent setup
- Should You Build a Loop? — stop signals and repeat-call caps in agent loops
- Claude Code Scheduled Tasks — where recurring Claude Code jobs live and can be ended