In Plain Sight - 16
In Plain Sight is a weekly blog from Rhys that we will publish every Thursday/Friday. Each edition will be focused on a thorny topic within investing, startups, learning and work and policy.
In Plain Sight is partially a personal attempt to think independently and to avoid leaning too much on AI for answers to our questions. Will we use AI to edit and polish the text? Yes. But more importantly, will we come up with all of the ideas and analysis? Yes.
Here follows Edition 16 - 'Unexpected item in the bagging area'
------
Anyone who has used a supermarket self-checkout knows the scene. A row of tills, each with a small light on top. When things go well the lights stay green and customers scan, pack and leave. When something goes wrong a light turns red, and an attendant walks over, taps a screen and sends the customer on their way.
Increasingly, that attendant is all of us.
As AI agents take on more of the drafting, researching, coding and summarising, a lot of knowledge workers are finding that their day has a familiar shape. We set several agents going and then spend the rest of the day walking between red lights.
That can feel like a demotion. We’d argue it’s a different job, and a harder one than it looks.
Good self-checkout attendants are doing a lot. They read the whole row at once. They can tell which customer is about to struggle before the light changes. They know which errors the machine will clear by itself and which need them. They calm people down. They do this across six or eight tills while also stacking baskets and answering questions about where the batteries are.
What they’ve lost is the conversation. A cashier used to spend a few minutes with each customer: a chat about the weather, a regular asking after the family. The attendant’s interactions are shorter and more transactional. Worse, they only ever arrive at the moment of frustration. Nobody calls them over when things are going well. They are, by design, the person standing between the customer and the exit and the customer’s mood usually reflects that.
That shift applies to us too. Orchestrating agents means many short interventions in place of a few long ones. Every red light is an interruption, and the default feeling on both sides is impatience with whatever is in the way.
Not all red lights are the same
The red lights come in two kinds.
The first is friction. An unexpected item in the bagging area. A barcode that won’t scan. These are system failures, and they get rarer as the technology improves.
The second is authority. Someone buying wine needs a person to confirm they’re old enough. This isn’t a failure. The system is built to stop and hand a judgement call to someone accountable.
From the customer’s side the two feel identical: something is stopping me. For the person orchestrating, telling them apart is the whole game.
With AI agents, friction looks like rephrasing a prompt, fixing a broken integration, or pasting in context the agent should already have had. The tools will keep removing these. Authority looks like deciding whether this email should go out, whether this analysis is right enough to act on, or whether this is the deal we want to do. Those decisions won’t disappear. As agents do more, each one carries more weight.
A career built on clearing friction is exposed. A career built on exercising authority well is not.
The teacher with thirty screens
The supermarket only goes so far. The customers at the tills are buying broadly similar things, and the attendant’s fixes are mostly interchangeable.
A better comparison might be a teacher in a classroom where every child is working on a device. Each one is on their own path, at their own pace, often on different material entirely. Hands go up. They can’t all be helped at once. And when the teacher kneels beside a desk, the context is entirely individual, because that’s the point of personalised learning. This child is stuck on fractions because of a misconception from last term. That child isn’t stuck at all, just bored. The one in the corner hasn’t put their hand up and needs help more than anyone.
The mental effort involved is considerable. The teacher has to pick up each child’s context in seconds, make a useful intervention, put that context down and pick up the next. Teachers have always done this. Technology has multiplied it.
Orchestrating agents is closer to the classroom than the supermarket. Each agent has its own thread, its own history and its own way of going wrong. The skill is loading that context quickly, helping well and moving on without losing track of the rest.
Several professions have lived with this for years, and they offer lessons.
Ward nurses answer call bells and monitor alarms across a whole ward. Healthcare has studied what happens when there are too many alarms: people start tuning them out. The term is “alarm fatigue,” and it’s a useful warning for anyone approving their twentieth agent output of the day.
Toyota’s production lines use an andon cord. Any worker can pull it to stop the line, and a team leader comes over. The culture treats a red light as useful information rather than a failure, and recurring stops are fixed at the root. As we wrote in Fix the Pothole, the goal is to make the problem disappear, not just easier to live with.
Grandmasters giving simultaneous exhibitions walk from board to board and take in each position in seconds. The best ones save their thinking for the few boards that need it.
When orchestration gets out of hand
There’s a trap here. Starting an agent is cheap, so it’s easy to start too many. Tasks are poorly scoped, agents overlap, one waits on another, and the number of red lights grows faster than anyone can clear them. It’s like a supermarket opening twenty self-checkouts with one attendant. On paper there’s more capacity. In practice the queue gets longer.
It’s possible to be extremely busy orchestrating and still produce less than you would have done alone.
That’s why we think an orchestration audit is becoming a useful habit. Every so often, step back and ask:
- Which agents produced most of the value, and which just produced activity?
- Which red lights keep coming back? Those are friction to fix at the root, not to keep clearing.
- Which tasks never needed an agent at all?
- Which red lights could safely turn green, with authority you could delegate? And which green lights should be red, where things are getting waved through?
- How many tills can you realistically watch before your judgement starts to slip?
The aim is fewer, better-designed streams of work, not more of them.
Done well, orchestration is a specific set of skills: breaking work into pieces an agent can run with, briefing clearly, knowing where errors hide, checking output without redoing it, switching context quickly and knowing when to leave things alone. Good managers, editors, producers, nurses and teachers already have these skills. AI is widening the group of people who need them, and it’s getting earlier in their careers.
If orchestration becomes nothing but responding to alerts, work gets busier but thinner, with less of the human connection that made many jobs rewarding. There’s an apprenticeship problem: people learn judgement by doing the work, and it isn’t clear where it comes from if juniors only supervise. And alarm fatigue applies here too. Tired orchestrators wave things through.

.png)

.png)
