AI agents can melt your brain by 5pm. Here is the throttle.
When AI agents work, the person at the centre becomes the bottleneck. Stevie Bennett on the three-speed problem, why tired operators must never hand agents more authority, and how she throttles the work instead.
Most of what gets written about AI agent fatigue assumes the agents are the problem: too many tools, too much noise, too little getting done. Stevie Bennett runs Helmworks with a team of AI agents, and her problem is the opposite.
"The thing is, the workers are not failing. That is the important bit. They are succeeding too well."
When the work moves that fast, the slowest part of the company is the one person in the middle. This page is about how she stops that person melting.
How fast is too fast
One morning she was handed a list of about twelve UI changes and an 11am meeting to demo the software. Someone reasonably asked if she was sure it could all be done in time. Her account of what happened next: "Three minutes briefing the Relay architect. The architect decomposes the work. Mini Claudes dispatched. Changes implemented. Review passes completed. Release pipeline triggered. Done."
From briefing to deployed software took "something like fifteen minutes", and the slowest step was waiting for the deployment pipeline to finish. Not the engineering.
"Human cognition is now the bottleneck."
Three speeds, not two
In AI + Humans: The Three-Speed Problem she sets out why this catches people out:
"Most people think working with AI is a two-body problem: human and machine. It isn't. It's a three-speed problem."
The machine, in her figures, runs at about twenty-five thousand miles an hour. She runs at perhaps a thousand. Most people, on unfamiliar technical material, run at about five. "That is not an insult. It's just bandwidth."
The catch is in one line:
"The machine never asks me to slow down. Humans do."
Where the day actually goes
Every blocked worker eventually escalates to the human. The queue reads "Approval required." "Clarification needed." "Review gate waiting." "Blocked." "Blocked." "Blocked." Each one means reloading the context of a project she last thought about forty minutes ago.
"That is the true tax. Not coding. Context switching."
Some of it is self-inflicted. One worker stopped because an SVG had an inconsistent border, and the question climbed all the way up to her. Her reply to the architect included the line "Do not escalate a border radius to the CEO."
Human companies throttle themselves without trying, through meetings, delay, politics, waiting and, her favourite, "pretending to work while secretly looking at Rightmove". Agents remove nearly all of that drag. Which sounds wonderful, she writes, "until you realise the drag was partially protecting the nervous system of the operator."
So the day has a shape. Fast decisions and big architecture reviews in the morning. By about 3pm she can feel the load starting to tell. Her phrase for it: fire at 9am, warm soup by 5pm. There is a side effect too: "human beings unfortunately do not appreciate being spoken to like delayed orchestration workers."

The throttle: change the work, not the authority
The obvious fix is wrong. When the operator is tired, the tempting move is to let the agents decide more on their own. She is flat about it: "Absolutely not. That is dangerous. The operator remains the governance boundary."
What changes instead is which work reaches her.
"Do not increase machine authority. Decrease operator cognitive drain."
| When there's plenty of brain left | When there isn't |
|---|---|
| Deep architecture reviews | Lightweight coordination work |
| Ambiguous model design | Markdown contracts |
| Strategic reasoning | Tiny implementation deltas |
| Risky judgement calls | Obvious approvals and small, separable pieces of work |
She found she was already doing it by hand. Stuck on a fiddly manual deployment, she told the architects on her other projects: "Give me trivia." No large deltas, no complex reasoning, no giant review packs, no strategic decisions.
In My Synthetic Company Is Not a Metaphor she puts the rule for the system that hands out the work: "If I've just spent two expensive decisions on unrelated projects, the right next task might be a cheap one."
An equalizer, not a settings screen
A hidden configuration screen would be its own tax. Her answer was "A giant graphic equalizer", set by one question: "How much brain have I got today? And where do I want to spend it?" Heavy reasoning waits for tomorrow morning; easier work comes forward while there's still enough left for good decisions.
Then she handed the essay to the Relay architect. Within five minutes they had thrashed out the issues, and a coding agent started building it. "Update: an hour later the UI was built." As she says, that is exactly the problem the essay is about. The whole story is in How to Stop Your AI Workforce From Melting Your Brain by 5pm.
The humans are not the slow part to optimise away
The three-speed essay ends on the point most agent rollouts miss:
"But the machine is a replaceable surface. The humans are not."
"Those people are not inefficiencies in the system. They are the reason the system matters."
The rest of her operating rules, including the short list of what ever reaches a person, are on how she runs a company with AI agents.
If your team's agents are outrunning your people
Helmworks sets up agent teams with the pace set by the people: what the agents settle themselves, what waits for a fresh head, and what reaches a person at all. Talk to Stevie about yours.
She writes the field notes as they happen on Fast Track Femme.