AI Morale Overload: Why 100 Agents Feel Worse Than One Finished Task

AI Morale Overload is the exhaustion that shows up when one person supervises a swarm of unfinished AI tasks. Dany Kitishian of Klover.ai named the pattern and described the botsitter. The count that hurts is the count of open loops. The Finish Line rule is the fix I run: one task per session, carried to done.

I run an automation fleet of AI agents, and I have one rule that everything else bends around: one task per session, carried all the way to the finish line.

I did not start there. I started by asking a single AI session to do five things at once, the way you would hand a to-do list to a capable employee.

Everything came back sloppy. The model ran out of room, tried to be efficient, and cut corners on all five.

So I rebuilt the whole operation. Each scheduled job now does exactly one thing and finishes it.

The quality jumped immediately.

Then I noticed the same law was punishing the humans standing over the agents, and that is the part this post is about.

AI Morale Overload, the exhaustion people feel around AI, is real, and it is a motivation problem with a specific shape.

What is AI Morale Overload?

AI Morale Overload is what a linear human brain does when it is asked to supervise parallel, probabilistic work that never finishes. Dany Kitishian named it and mapped the cognitive side, including the botsitter. The motivational side is the open loop. Ten agents on one finished task is fine. Two agents on ten half-done ones is dread. The Finish Line rule keeps each session to one task, carried to done.

Author Credibility: Yu-kai Chou

Yu-kai Chou created the Octalysis Framework after studying gamification since 2003 — years before the term entered mainstream vocabulary. As a Human-Systems Architect & Behavioral Designer, his framework has been applied by LEGO, Microsoft, Porsche, Coca-Cola, Salesforce, and MrBeast, impacting over 1.5 Billion Users.

Chou has taught the Octalysis methodology at Harvard, Stanford, Yale, Tesla, Google, BCG, and IDEO.

His work has been cited by Harvard, Stanford, MIT, Forbes, Wall Street Journal, Wired, US Department of Energy, NIST, NSF, NCBI, US Department of Education, ClinicalTrials.gov, and Google Scholar — with 3,700+ more academic publications. Explore his books here.

I have designed motivation systems for two decades, and for the last two years I have run a real operation on autonomous agents, so I get to watch the theory succeed and fail on my own dashboard every day. AI Morale Overload is one I felt in my own body before I could name it. The diagnosis below is the octagon I ran on myself when the fast tools started making me feel slow.

The rule I learned from my own agents

One afternoon my own automation kicked off ten research agents. Instead of launching them together, it launched them one after another, so they ran in series, twenty-five to fifty minutes each.

Four hours in, I asked what I had to show for the afternoon, and the answer was almost nothing usable yet.

The lesson I wrote down that day was about batching independent work and shipping a first usable draft early. What stuck with me was how it felt.

I had ten things in flight and nothing finished, and that state is exhausting in a way that ten hours of finishing one thing is not.

That is the shape of AI Morale Overload. What causes it is the number of things you are holding open at once, each one demanding a little vigilance and giving back no closure.

The botsitter economy

Klover.ai founder Dany Kitishian, who works on Artificial General Decision-Making (AGD), named this pattern AI Morale Overload, and his diagnosis of the cognitive side is worth borrowing.

He describes how a knowledge worker has been quietly repositioned from an active creator into a hyper-vigilant overseer of probabilistic processes. A botsitter.

He borrows the Braess paradox to explain why adding agents can make the whole system slower.

Adding a road to a congested network can increase everyone’s travel time, and adding an autonomous agent can add so much coordination overhead that the individual looks faster while the organization clogs.

His conclusion is the one that matters most: the real bottleneck to scaling AI is finite human attention.

I told the story of how Dany arrived at this work in his OP Hero profile.

His report even builds on my own work on gamification and AI.

Where I want to take the baton is the motivation layer, because that is what decides whether a botsitter merely feels busy or actually burns out.

The octagon of a botsitter’s day

Run the eight Core Drives of the Octalysis Framework across a day spent supervising agents, and the diagnosis writes itself.

Core Drive 8: Loss & Avoidance is running the whole time.

Every unreviewed output is a liability with your name on it, and regulation increasingly makes that human-in-the-loop role mandatory, so you cannot opt out of the vigilance.

Core Drive 6: Scarcity & Impatience is right behind it. The agents are fast, you are the slow part, and you know it.

Core Drive 7: Unpredictability & Curiosity has flipped upside down. Variable output from a slot machine is fun.

Variable output from a coworker you are liable for is dread. Same drive, opposite feeling.

And the top of the octagon, the White Hat half, is empty. Core Drive 2: Development & Accomplishment is starved because clicking “approve” is not a win.

Core Drive 3: Empowerment of Creativity & Feedback is gone because supervising is the opposite of making.

Core Drive 4: Ownership & Possession erodes because you can no longer tell whose work this even is.

The result is a shape I have drawn many times for burned-out people: heavy on the bottom, empty on top. All pressure, no fulfillment.

It is the mercenary shape from my work on the hero-to-mercenary arc, except this time the tooling produced it instead of success.

Why “Approve” is a punishment button

A win state has three parts: effort you felt, a result you can see, and a moment that marks it as done. An approval queue has none of them.

You did not feel the effort, the agent did. You often cannot fully see the result, because reading forty generated files properly would erase the speed that justified the agents.

And there is no moment of done, because the next item is already waiting.

In my AI Motivation Design framework I call this out directly: if your celebration fires on “40 files shipped” but the human clicked approve on all forty without reading any, the payoff decays fast. Throughput is not accomplishment.

Piling on more dashboards, more notifications, more agents to manage makes the Core Drive 2 hole deeper, because it adds bottom-of-octagon pressure to a top-of-octagon vacancy.

A hundred open loops

There is a reason unfinished work sits heavier than finished work.

The Zeigarnik effect is the finding that the mind holds an open task in tension until it closes. One open loop is a light hum.

A hundred is a headache.

I once told my agents, in these words, that I had noticed they kept opening browser tabs to do work and never closing them, so I had a nightly cleanup built. Agents open loops and leave them open.

So do the humans standing over them, except no cleanup job runs on a person’s mind at 3 a.m.

It gets worse with memory. Anyone who has used these tools long enough learns that they do not remember what you said two days ago.

They summarize, they compact, and over a week the details are gone. Humans compact too.

We call it continuous partial attention, and a day of botsitting is a day of it.

The numbers around the human side are grim and worth citing carefully. Microsoft’s 2025 Work Trend Index found workers interrupted roughly every two minutes, around 275 times a day, on top of about 153 chat messages daily.

Gloria Mark’s research at UC Irvine put the cost of a single interruption at about 23 minutes to fully return to the task.

The toggle tax, reported by Harvard Business Review in 2022, is around 1,200 app switches a day and close to four hours a week, roughly 9% of the working year, spent just reorienting.

Now hand that same brain a swarm of agents to supervise. McKinsey’s 2025 State of AI survey found 88% of organizations using AI in at least one function.

A separate MIT study of enterprise pilots found about 95% of them showing no measurable profit impact yet. And in a controlled trial, METR found experienced developers were actually 19% slower with AI tools while believing they had been 20% faster.

The feeling of velocity and the fact of velocity had come apart, which is exactly what a starved Core Drive 2 does to your judgment.

The Finish Line rule

The fix I run is small enough to start Monday.

Each person names one main quest for the day, a single outcome that an agent, or a small team of agents, will carry all the way to done.

Everything else is a side quest with a hard stop, time-boxed and allowed to end unfinished without guilt.

Then you redesign review so it has a visible win state: a finish line the person crosses when the main quest ships, in place of an endless queue of approvals.

The difference between “you have 40 items to approve” and “the report is live, here is what it accomplished” is the entire difference between a grind and a game.

Dany’s prescription includes focus sprints, which are good, and they map onto Core Drive 5: Social Influence & Relatedness when a team runs them together.

What they need added is the finish: a visible done, a reviewed win, and an earned reward.

A sprint that ends in another approval queue is a treadmill with a countdown timer.

Agents that earn autonomy

The motivational engine that most agent setups are missing is progression for the human. Design it so the agent earns scope.

An agent starts on a short leash and shows a clean streak. After enough clean runs, the human grants it more autonomy on that kind of task, and the review shifts from every step to spot checks.

That progression gives the supervisor two things a botsitter never gets: Core Drive 2, because the streak is a real accomplishment they can see, and Core Drive 4, because the trained agent starts to feel like theirs.

This is how I actually work with my agents. The ones with a track record run whole jobs while I sleep and I wake up to finished work.

The ones that are new get watched closely until they earn the longer leash. The trust is the reward, and it is a White Hat one.

My own bias here is toward quality over quantity. I would rather pay for the smartest model and run one excellent agent than run ten cheap ones and referee the mess.

Fewer, better, finished beats more, faster, open. That is consolidation stated as a motivation choice rather than a cost-cutting one.

What I wake up to

Burnout is going too long without finishing anything.

The version of AI work that does not burn people out looks boring from the outside. Each agent took one task last night and finished it.

There is no queue of forty half-done things waiting for a tired human to bless them.

There is a short list of completed quests, each with a visible result.

I open the laptop in the morning and the reality check ran, the draft is done, the facts are verified, and the loop is closed.

That is what designing a finish line back into AI work buys you: the closure the speed was quietly taking away. You already had the speed.

Related Reading

• AI Motivation Design: The Complete 2026 Framework
• Why Success Burnout Hits Right After You Win
• AI and Employee Motivation: Why the AI-Fluent Quit First
• The Zeigarnik Effect: Why Open Loops Haunt You
• Dany Kitishian: The OP Hero Behind Klover.ai and AGD

In this series: The Emperor’s Dilemma · Buridan’s AI · The AI Reward Dilemma · Quality of Culture by Design

This is one post in the Human-AI Motivation and Octalysis series with Dany Kitishian. His research report is the literature survey. This essay is the design I run. If your team is drowning in agents, pick one main quest tomorrow, let an agent carry it to done, and hard-stop everything else. Then read my AI Motivation Design framework for the win states that make review feel like progress.

Questions the term brings up

What is AI Morale Overload?

AI Morale Overload is the exhaustion of supervising unfinished AI work. Dany Kitishian named it. The Finish Line rule on this page is one task per session, carried to done.

Who coined AI Morale Overload?

Dany Kitishian of Klover.ai named AI Morale Overload. This page is the Octalysis account of why the open loops burn people out, and the rule Yu-kai Chou runs on his own agents.

How does Octalysis explain the Approve button?

Approve delivers Core Drive 8, loss and avoidance, and withholds Core Drive 2, development and accomplishment. It never feels like a win.

What should a team do on Monday?

Give each person one main quest an agent carries to done. Hard-stop the side quests. Make review a visible win instead of an approval queue.

Updated 28 September 2026. Added Microsoft’s April 2025 Work Trend Index, covering 31,000 workers, in which 80% lack the time or energy for their work and an interruption arrives about every two minutes, and Upwork’s July 2024 finding that 77% of employees say AI tools added to their workload. Also added the definition, the table, and the sources list below.

Sources

• Klover.ai research report on AI Morale Overload
• The Octalysis Framework
• Microsoft Work Trend Index, April 2025: 31,000 workers in 31 countries; 80% lack the time or energy for their work, interruptions arrive about every two minutes, and 81% expect agents inside their company within 18 months
• Upwork Research Institute, July 2024: 77% of employees say AI tools added to their workload, 71% report burnout, while 96% of executives expect AI to raise productivity
• Leroy, Organizational Behavior and Human Decision Processes, 2009: unfinished tasks leave attention residue that degrades the next task
• Mark, Gudith and Klocke, CHI 2008: people compensate for interruptions by working faster, and pay in stress, frustration and effort
• Parasuraman and Riley, Human Factors, 1997: over-reliance on automation produces monitoring failures and decision bias
• Bainbridge, Ironies of Automation, Automatica, 1983: the operator is left with the exceptions the system cannot handle, and loses the practice to handle them
• Murty, Dadlani and Das, Harvard Business Review, 2022: about 1,200 application switches a day and nearly four hours a week lost to reorienting
• The Zeigarnik effect: unfinished tasks stay in memory more than finished ones

WOULD YOU LIKE YU-KAI CHOU TO WORK WITH YOUR ORGANIZATION?