Guest: Jon Berger, technologist, management and leadership expert and author of What Happens When You Are Not In The Room?
Watch on YouTube
Listen on Spotify
Listen on Apple Podcasts
Read shownotes & transcript below
Asynchronous & Unreliable | Anne Currie with Jon Berger
What do steam engines, fighter pilots, Japanese manufacturing, software development and AI agents have in common? Feedback loops.
In this episode, Anne is joined face-to-face for the first time by returning guest (and husband!) Jon Berger, author of What Happens When You Are Not in the Room?, to explore one of his favourite subjects: feedback loops and why they're fundamental to building better systems and better organisations.
Starting with the mechanical governors that regulated early steam engines, Anne and Jon trace the evolution of feedback loops through W. Edwards Deming's revolutionary approach to manufacturing quality, John Boyd's OODA loop, and the rise of iterative software development.
They discuss why feedback loops are deceptively difficult to implement, why even experienced engineers struggle to abandon an overreliance on upfront planning - good upfront planning remains valuable but cannot anticipate everything - and why successful systems need ongoing observation, learning and adaptation rather than a one-and-done approach. They also explore how continuous improvement changes not just processes but the autonomy of the people operating them.
Finally, they return to the question that inspired the conversation: can feedback loops help AI agents become genuinely autonomous, rather than simply executing increasingly elaborate prompts?
What is a feedback loop? Using the output of a system to assess and improve its future behaviour.
Steam engines and mechanical governors: How automatic feedback transformed early industrial machinery.
Feedback loops in software: Autoscaling, why scaling down is often neglected, and how to recognise an incomplete loop.
Deming and PDSA: Plan, Do, Study, Act — the origins of continuous quality improvement and its influence on Japanese manufacturing.
Boyd and OODA: Observe, Orient, Decide, Act — why separating observation from interpretation is so valuable for teams.
From waterfall to iteration: How CI/CD, testing and continuous delivery support an entirely different approach to software engineering.
Why loops are hard: The persistent temptation to want to design everything perfectly upfront, even when requirements and circumstances inevitably change.
Management and autonomy: How shared goals and effective feedback loops allow teams to improve their work without constantly involving their manager.
AI agents and feedback: Why defining what good looks like may be more powerful than crafting the perfect prompt.
Jon Berger — What Happens When You Are Not in the Room? A book about management, leadership and building effective, autonomous teams. Paperback available now, with the Kindle edition scheduled for 1 December 2026.
W. Edwards Deming — PDSA and continuous quality improvement.
John Boyd — The OODA loop and decision-making.
Toyota Production System — Continuous improvement in manufacturing.
Eliyahu M. Goldratt — The Goal — A novel about manufacturing, systems thinking and organisational improvement.
Yanqing Cheng — Previous podcast guest exploring feedback loops and AI agents.
Holly Cummins — Previous podcast guest mentioned in the discussion of autoscaling and software efficiency.
Asynchronous & Unreliable: Conversations at the edge of software.
Hosted by Anne Currie, exploring distributed systems, software engineering, resilience, AI and the people building our technological future.
Website: https://www.asynchronousunreliable.com/
Anne: Hello and welcome to this week's episode of Asynchronous and Unreliable, a weekly podcast where we discuss the most interesting ideas and concepts in tech. I'm your host, Anne Currie, author of Building Green Software from O'Reilly and the science-fiction Panopticon series.
Today I'm going to do something a little bit different. I've mentioned before that my regular guest, Jon Berger, also happens to be my husband, and is generally cached locally! Today, for the first time, we're actually going to do a face-to-face episode, so you'll have to forgive us for struggling a little with this, because neither of us is an actual TV presenter. Welcome, Jon. Welcome back to the show.
Jon: Thanks. Always good to be back.
Anne: I've got two motives for having you on today. One is that I've spoken a few times recently with guests, particularly Yanqing Cheng, about feedback loops. I know that's something you're really interested in. You wrote a lot about it in your book, What Happens When You Are Not in the Room?
The other thing I want you to talk about is the Kindle version of your book and when it's coming out. Viewers will be able to see that I'm holding up the physical version. That came out about a month ago now — two months ago?
Jon: Yes, it came out in September. The Kindle version is going to come out on the first of December.
Anne: Which is good. I'm looking forward to that.
Jon: And there will be an introductory price.
Anne: So it'll also be cheap, which is nice. Which makes sense, doesn't it? It's always slightly galling when the Kindle version of a book is the same price as a physical version. It always feels like it should be a lot cheaper.
But we're going to talk a little about What Happens When You Are Not in the Room? What I really want to talk about is feedback loops.
Yanqing was trying to make her AI chatbots understand and implement feedback loops. In the episode she recorded with me, she hadn't yet managed to make that work, but she's already told me she's now been able to do it with more modern versions of the AI. So I'll get her back on to talk about that.
It occurred to me while I was talking to her that I really needed to get her to explain what a feedback loop was, why it was a good idea, where it came from, and why it was a useful management tool when she was trying to train AIs to be better at management. But I didn't ask her. I know I can ask you, because you wrote all about it in your book. So tell me: what is a feedback loop?
Jon: You have something you're trying to improve, or something you're trying to stabilise, where you want to continually look at the output of what you're achieving and say, “That's right; that's not right. How can I make it better?”
So you have an input: the loop starts, something kicks off. You're doing some work, or you're doing something mechanical, and you have an output. If you don't have a loop, input goes to output and that's that. You never look at it again.
But if you want a loop, you look at that output and go, “Is that what I wanted?” That's normally called the signal. What is it that I'm seeing in this output that's different from ideal, for example, or can be improved? You take that signal, turn it into feedback, feed it back into the process at the start of the loop, and change what you do so you're more successful the next time you go through. Then you just keep going through the loop.
Originally these loops were described in the context of mechanical governors. Steam engines, for example, had these governors. I don't know whether I can describe this, but I'll give it a go. You would have two steel balls that would spin round. You'll see these in pictures of early machines. They'd be attached in a diamond shape, with the steel balls at the left and right sides, whizzing round.
As the machine went faster and faster, centrifugal force would push out those balls so the diamond would squash down. That would drive a piston or a lever, which slowed the machine down. As it slowed down, the balls would come in and the diamond would elongate vertically. That's feedback: the speed at which the thing is going round controls the speed at which the engine is running.
Anne: Which is a really good example, and I like it. But that's an example where a feedback loop has already been automated into your system. With management, what you're talking about in your book isn't a situation where there's already a feedback loop automated into your system. For most companies, you don't have that yet.
Jon: The more automatic the better, and what Yanqing wants is fully automatic. But loops are used in all sorts of ways, and they are used automatically in systems as well. If you have a system that's automatically scaling up and scaling down based on workload, you've got something monitoring the load, saying, “Wait a minute, I'm a bit overloaded; spin up another machine, spin up another VM,” or “I'm now actually pretty flat.” That's a loop.
Anne: So, in the tech industry, autoscaling, effectively. But isn't the interesting thing about the tech industry that there are loads of counterexamples as well? We've gone off on one, because all these very detailed examples of feedback loops aren't necessarily great examples.
For example, autoscaling — I talk about that a lot in Building Green Software, and one of my other podcast guests, Holly Cummins, talks about it quite a lot as well. One of the problems with autoscaling is that it's not a very good feedback loop. It's only one way. People tend to scale up, but they never scale down, because it's much safer to overprovision than underprovision. So people automate the “this needs to be bigger” part, but they don't complete the loop with “and now this can be smaller again.”
I would say we've leapt ahead a little bit here, which is fine. We're allowed to leap ahead!
Jon: Well, what you're saying is it's possible to implement loops badly. And it is. That's an example. If you only ever move in one direction, and you're never looking at the signal that says, “I've got vastly more deployed resources than I need,” then either you're missing another loop that needs to run, or you've poorly implemented that loop.
It's very easy to poorly implement loops, and you'll probably get it wrong when you first do it. One of the things you want to do is have a loop to study your loops.
Anne: Now we're galloping massively ahead again. I'm going to roll it all the way back. A loop is: you do something, you look to see whether it works or not, you make a change, and then you do it again and see whether it got better or worse. Fundamentally, is that what a loop is?
Jon: There are actually a couple of different types of loop that I talk about in the book. What you've described is a classic feedback loop. Exactly that.
I do have quite a few loops, and quite a lot of bits in the book about loops, because I like loops. When I first told people I was writing a book on management, several of them said, “Is it about loops?” It's like, “It is about loops, but it's not only about loops. It's about other things too.”
In management, a lot of people will have heard of the OODA loop or the PDSA loop. PDSA is clearly a feedback loop: Plan, Do, Study and Act. You plan to do something, then you do that thing. If you didn't go any further, there's no loop. But you study what you've created and say, “Is that what I wanted?”
These were developed by W. Edwards Deming as part of measuring and improving quality. He'd look at what had just been done: “Is that car up to quality? Is the thing my machine has just built at the quality level I want?” Let's do some testing, preferably automatically. If it isn't good enough, you find that in your study, then act — or adjust — so that the next time around the loop, the output is of higher quality.
Anne: So Deming was an early pioneer of organisational and manufacturing studies, and he did a lot of his work during the First World War —
Jon: Second World War.
Anne: Second, sorry. Second World War. It was about getting your tanks and aeroplanes out faster and to a higher quality.
Jon: Yes, but while much of his research was done during the war, the point at which his work really became famous was immediately after the Second World War, particularly as part of rebuilding Japan and getting its industry on track. That's where people started saying, “Wait a minute, this guy seems to know what he's talking about.”
Anne: And that led to the Toyota Production System.
Jon: Yes, the famous Toyota Production System: just-in-time and all that management thinking that revolutionised the way things are produced at scale. But it needed his loop to control the quality so that you could operate at scale and change things without breaking things. That was his PDSA loop: Plan, Do, Study, Act — or Adjust, as some people say.
Anne: So we've talked a lot about PDSA loops, but they're not the most famous loop at the moment, at least in the tech industry. I don't know whether that's more widely held. The famous one isn't Deming's PDSA loop; it's Colonel Boyd's OODA loop. Tell us a little about OODA loops. How are they the same and how are they different? You cover both in your book.
Jon: Colonel Boyd was a US Air Force fighter pilot in the Vietnam War. Coming out of that, he came up with the OODA loop, which probably is the most famous loop. OODA stands for Observe, Orient, Decide and Act.
You start by observing what's going on, then you orient yourself. Observe is objective, dispassionate: “This is what I see.” Orient is your subjective interpretation: “In my context, what does that mean?” I really like the difference between those two. That's why I personally love OODA: separating those things out.
Then you decide: “Based on what I'm seeing and what I think about it, this is what I want to do.” Then you act, and see what impact that's had.
When Boyd created it, the idea was that if you're in a dogfight you want to go round that loop as quickly as possible. The pilot who goes round the loop fastest wins, because they're making the right decisions in context, very quickly. I have no idea whether that's actually useful in a dogfight, but I think it's incredibly valuable in a business context. It's certainly used much more widely than the tech industry, which is probably why it's so famous.
For me, the biggest value is the difference between Observe and Orient, because it means you can go through that loop with other people. Generally speaking, we're not in dogfights. We're not under as much time pressure. Being able to separate Observe from Orient means you can go through the loop with the rest of your team, with other people who care about getting a better result.
You might all see the same thing and agree, “Yes, that is what we observed,” but different people have different interpretations of what it means. That's super valuable: “Why are we interpreting this differently?” Using that as a tool of alignment can be really helpful. Then you decide. You're separating those things out instead of just going, “Let's do a thing, let's do a thing, let's do a thing.”
It's almost — not entirely, but almost — as though one of the values of using it in business is the opposite of what Boyd came up with. He came up with a loop you should go round as quickly as humanly possible, but in business its value is almost that it slows you down. Instead of “Let's do this, let's do this, it worked, let's do this,” you ask, “What have we actually seen? What do we think happened? What are we going to do about it?” And then you do it. Different people can say, “Wait a minute,” at different phases of the loop.
Anne: At the beginning you were talking about governors, an example of physical feedback that revolutionised another industry: steam engines. That was a long time before Deming. You had governors, and then 150 years later, or whatever, you had Deming's PDSA. That was about how to do better at producing planes, tanks and guns that work.
His approach was to say, “Look at what we've been doing.” With loops you go, “Well, that's a no-brainer. Of course you'd look at the results of your systems and decide whether they're good enough or how to make them better.” But it wasn't obvious, was it?
Jon: Well, that wasn't really how people thought about manufacturing back then. Going back to Ford, probably the most famous person in terms of revolutionising large-scale manufacturing with cars: you'd put an enormous amount of effort into setting up your production line. You'd try really hard to ensure up front, before any cars rolled off the line, that it would create the quality you wanted.
Then at some point later you'd create a new production line, probably for a new model of car, and it would be different. Whereas what Deming was trying to do was look at each thing that rolled off the production line — in an ideal world. In practice, you couldn't look at each thing because you didn't have that level of automation when building physical things. So you'd look at every thousandth item, say, and test whether it was within tolerances. If it wasn't, you'd do something about it.
A whole bunch of other interesting things come out of that: andon cords and production systems. You've got that ability to check as you go along: “Is the machine doing what we want? Can we improve it? Can we create better tolerances?”
And it changes the attitude of everyone in the factory. They're thinking, “We're not just turning the handle. It's not just that the machine has been set up, the boffins have done their job and now we have to feed things in at one end and take them out of the other.” It's everyone's job to maintain and improve quality continually throughout the manufacturing lifecycle of that machine.
Anne: Yes, and if you can improve the machine as you go, people's lives get even better and the output gets better. One really interesting thing is how hard a cultural shift that is.
You and I started in the tech industry 30 years ago, and at that time it was the pre-Deming way. We did waterfall. We did loads and loads of thinking up front. All our brains were going like crazy, planning and planning. I kind of liked that! I really liked that.
Even now, it's quite recent that people have moved to a more iterative, fast-flow way of approaching things. I wouldn't even necessarily call it agile. It's iteration: doing lots of small things. CI/CD is the key tool here. You need something like that. It's all about testing. For Deming it was about quality: how do you improve that? CI/CD is a key tool there.
Jon: It's not quite the same — in fact, it's very much not quite the same. In physical manufacturing, you were building a machine once that would go away and generate a million other things: “Build me a million cars, build me a million washers.” It was worth putting a huge amount of effort in up front. You can see why people thought it was all about that upfront effort.
But loops aren't saying, “Don't bother planning or putting effort into creating it up front.” They're saying, “Don't think that's the end of it.” That's one of the things I like most about all these loops: they don't end.
Software development doesn't end until no one cares about your product any more. The point where you ship your product is really the start. That's when your customers first say, “That's exactly what I wanted, apart from this,” or “It's not at all what I wanted.” Any successful software product spends vastly longer in maintenance, support and engaging with customers than in initial development.
People spend all their time worrying about making the initial development great. But if you think of it as a loop, you're continually improving the product. Test-driven development, behaviour-driven development and the various other DDs try to move a little in that direction.
It's different from manufacturing because you're still really only building one product. You're not saying, “Now the code works, spin up a million copies of it.” You're sort of doing that, but that's not the machine doing it.
If you're in an environment where things change quickly — or really, change at all — those loops help enormously. You don't say, “We've done a thing. Finished.” You say, “We've done a thing. Now we do another thing, and another.” You've got the tooling, such as CI/CD, and the attitude of your team: we're constantly changing and improving. It's the expectation.
That's why I've always thought loops aren't just about the part of software development where you have a requirement and a product that meets it. They're about everything else in how you work. How are you developing your people? How are your teams operating? Are they happy? Are they productive? Are they doing what they think they should be doing? What's your customer-support process like? Everything that's part of a business can be improved. And a loop is a brilliant way of doing that.
Anne: It's interesting. You talk about it and think, “God, it's a no-brainer. It's so much better. It completely revolutionised everything.” But it's worth constantly reminding ourselves that everybody wants to revert to the single big-shot delivery. Everyone wants to go down the waterfall in the end! Even when they don't, they build smaller waterfalls rather than continuous feedback.
In The Goal, by Eli Goldratt — this absolute blockbuster management book from the eighties, if such a thing existed then; I don't think it will ever exist again — he's describing that Toyota Production System kind of Deming thinking. Ironically, it was created by an American, moved to Japan and refined there, and then Goldratt was attempting to bring it back to the US.
I enjoyed the follow-up — I think it was probably the final version of his book before he died, sometime in the nineties — saying that the interesting thing is that even in Japan people don't all follow the Toyota Production System. Everybody knows about it, knows it's great and wants to do it, but hardly anyone can successfully pull it off. It's such a complete mindset shift from thinking that you're clever enough to work it all out up front. All engineers think that, whether manufacturing or software engineers.
And I'm talking about you and me here. We're both terrible for that sort of thing. We both think we're clever enough to think it through up front. But you have to accept that it's not about being clever or not. There are just so many unpredictable things. The world is very unpredictable, which is one reason it's hard.
Jon: But it's actually got nothing to do with how clever you are. That's the point. The fallacy isn't that people think they're cleverer than they are. The fallacy is that cleverness is even relevant, right?
Why are loops useful? Because they change your attitude, as I've just said, and because they allow you to change. Even if you are smart enough to think in advance of exactly how a motor car or a piece of software should be built, you can't predict every future demand on that thing: what your customers want, what your executive team wants, what the government allows you to do. The idea that your product should never change after that, or that you can predict all those changes, is crazy.
Say you create this brilliant product, you've thought hard about it, done a knock-out job of making it, and then a million people buy it. Still brilliant. Then the world changes. Do you want someone else to come along and do a slightly better job because they're more aware of the context? Or do you want to keep moving quickly to stay ahead of the changing requirements?
But it's not just that that makes loops powerful.
Jon: One of the reasons loops are in this book is that loops allow the manager to get out of the room. Once you've created a loop and people can operate it, they can improve things without you needing to be there with your giant brain telling them what to do.
While you're asleep, on holiday, or doing a thousand other things, they can see what needs improving. They can see, “We're not achieving what good looks like. How do we do that?” Or, “We are achieving what good looks like. How do we achieve what great looks like?” Running those loops allows them to do that irrespective of what you're doing.
One reason for that is you've probably spent some time discussing what “good” looks like. Getting that into the open and making it a shared vision is the key to allowing people to operate independently, go fast, and use their own expertise, awareness and understanding of the situation, whether you have that understanding or not.
Anne: Yes. This also reminds me why Yanqing is so keen on defining loops for AI agents. She wants to go to sleep and have an AI agent loop round and iterate towards a defined goal while she's asleep — or, let's face it, operating on a completely different timescale from her AI agents.
Jon: Absolutely. She wants them to improve things autonomously, as you'd want a good employee to do. Not that long ago there was a common fallacy, which we can very much laugh about now, that it was all about prompt engineering. Getting AI to do the right thing was all about crafting the perfect prompt so it could go away and do that thing.
No. What does good look like? If you can define that and allow an agent, or a team of agents, to iterate against it, they can do that. If you've got to write the perfect prompt up front, the only way you're going to make it successful is with another loop outside it that goes, “Your prompt wasn't good enough. How about this?” Then you're doing all the work.
Anne: Yes. And you're the bottleneck. You've created a bottleneck; you haven't actually delegated.
Anne: I think this has been an excellent episode. We've talked about quite a lot of stuff. I want you to push your book again!
Jon: Here is the paperback copy of What Happens When You Are Not in the Room? For listeners, I'm holding up a copy. That's already available to purchase on Amazon. If you want the Kindle version, it comes out on the first of December at a special introductory price.
Anne: Hopefully I'll catch you again on next week's episode of Asynchronous and Unreliable. I do run into quite a lot of actual, physical people who listen to or watch this podcast, and I usually get all my feedback from them. My feedback loop is going to the pub or for a coffee with somebody who says, “I watched your podcast,” and quite often they end up being guests.
Hopefully those people will let me know whether this is a good format or not.
Jon: Yeah.
Anne: Thank you very much for listening to Asynchronous and Unreliable. I'll catch you again next week. Thank you very much.