Measuring Inner-Loop Agility

When I teach teams and managers about the feedback loops of software development, I try to stress the two most important loops – the ones that define agility.

Releases are where the value gets delivered, and – more importantly – where the end user feedback starts, so we can learn what works and what doesn’t and adapt for the next release.

The sooner we can release, the sooner we can start learning and adapting. So agile teams release frequently.

But frequency of releases doesn’t really define agility. I see teams who release every day or every week, but feature requests still take months to get into production.

That feature-to-production lead time is our best measure of how responsive to change we’re really being. How soon can we adapt to customer feedback?

For a portfolio of software products, a client of mine plotted average feature-to-production lead times against the average time it took to build and test the product.

We see a correlation between that feature-to-production lead time and the innermost loop of software delivery – build & test time.

Of course, this is a small data set, and all the usual caveats about “lies, damned lies and statistics” apply (I would love to do a bigger study, if anyone’s interested in participating).

But I’ve seen this distribution multiple times, and experienced it – and observed many, many teams experiencing it – in the field.

Products with slow build & test cycles tend to have much older backlogs. Indeed, backlogs themselves are a sign of slow lead times. I explained the causal mechanism for this in a previous post about Inner-Loop Agility. When we want to optimise nested loops, we get the biggest improvements in overall cycle time by focusing on the innermost loop.

Now, here’s the thing: everything that goes on between releases is really just guesswork. The magic happens when real end users get real working software, and we get to see how good our guesses were, and make more educated guesses in the next release cycle. We learn our way to value.

That’s why Inner-Loop Agility is so important, and why I’ve chosen to focus entirely on it as a trainer, coach and consultant. I can’t guarantee that you’re building the right thing (you almost certainly aren’t, no matter how well you plan), but I can offer you more throws of the dice.

Code, Meet World #NoBacklogs

Do we dare to imagine Agile without backlogs? How would that work? How would we know what to build in this iteration? How would we know what’s coming in future iterations?

Since getting into Agile Software Development (and it’s precursors, like DSDM and Extreme Programming), I’ve gradually become convinced that most of us have been doing it fundamentally wrong.

It’s a question of emphasis. What I see is thousands of teams working their way through a product plan. They deliver increments of the plan every week or three, and the illusion of being driven by feedback instead of by the plan is created by showing each increment to a “customer” and asking “Waddaya think?”

In the mid-90s, I worked as the sole developer on a project where my project managers – two of them! – would make me update their detailed plan every week. It was all about delivering the plan, and every time the plan bumped into reality, the whole 6-month plan had to be updated all over again.

This was most definitely plan-driven software development. We planned, planned, and planned again. And at no point did anybody suggest that maybe spending 1 day a week updating a detailed long-term plan might be wasted effort.

Inspired by that experience, I investigated alternative approaches to planning that could work in the real world. And I discovered one in a book by Tom Gilb called Principles of Software Engineering Management. Tom described an approach to planning that chimed with my own experience of how real projects worked in the real world.

It’s a question of complexity. Weather is very, very complicated. And this makes weather notoriously difficult to predict in detail. The further ahead we look, the less accurate our predictions tend to be. What will the weather be like tomorrow? We can take a pretty good guess with the latest forecasting models. What will the weather be like the day after tomorrow? We can take a guess, but it’s less likely to be accurate. What will the weather be like 6 weeks next Tuesday from now? Any detailed prediction is very likely to be wrong. That’s an inherent property of complex systems: they’re unpredictable in the long term.

Software development is also complex. Complex in the code, its thousands of component parts, and the interactions between them. Complex in the teams, which are biological systems. Complex in how the software will interact with the real world. There are almost always things we didn’t think of.

So the idea that we can predict what features a software system will need in detail, looking months – or even just weeks ahead – seemed a nonsense to me.

But, although complex systems can be inherently unpredictable in detail, they tend to be – unless they’re completely chaotic – roughly predictable in general terms.

We can’t tell you what the exact temperature will be outside the Dog & Fox in Wimbledon Village at 11:33am on August 13th 2025, but we can pretty confidently predict that it will not be snowing.

And we can’t be confident that we’ll definitely need a button marked “Sort A-Z” on a web page titled “Contacts” that displays an HTML table of names and addresses, to be delivered 3 months from now in the 12th increment of a React web application. But we can be confident that users will need to find an address to send their Christmas card to.

The further we look into the future, the less detailed our predictions need to become if they are to be useful in providing long-term direction. And they need to be less detailed to avoid the burden of continually updating a detailed plan that we know is going to change anyway.

This was a game-changer for me. I realised that plans are perishable goods. Plans rot. Curating a detailed 6-month plan, to me, was like buying a 6-month supply of tomatoes. You’ll be back at the supermarket within a fortnight.

I also realised that you’ve gotta eat those tomatoes before they go bad. It’s the only way to know if they’re good tomatoes. Features delivered months after they were conceived are likely rotten – full of untested assumptions piled on top of untested assumptions about what the end users really need. In software, details are very highly biodegradable.

So we need to test our features before they go off. And the best place to test them is in the real world, with real end users, doing real work. Until our code meets the real world, it’s all just guesswork.

Of course, in some domains, releasing software into production every week – or every day even – is neither practical nor desirable. I wouldn’t necessarily recommend it for a nuclear power station, for example.

And in these situations where releases create high risk, or high disruption to end users, we can craft simulated release environments where real end users can try the software in an as-real-as-we-can-make-it world.

If detailed plans are only likely to survive until the next release, and if the next release should be at most a couple of weeks away, then arguably we should only plan in detail – i.e., at the level of features – up to the next release.

Beyond that, we should consider general goals instead of features. In each iteration, we ask “What would be the simplest set of features that might achieve this goal?” If the feature set is too large to fit in that iteration, we can break the goal down. We build that feature set, and release for end user testing in the real (or simu-real) world to see if it works.

Chances are, it won’t work. It might be close, but usually there’s no cigar. So we learn from that iteration and feed the lessons back in to the next. Maybe extra features are required. Maybe features need tweaking. Maybe we’re barking up the wrong tree and need to completely rethink.

Each iteration is therefore about achieving a goal (or a sub-goal), not about delivering a set of features. And the output of each release is not features, but what we learn from watching real end users try to achieve their goal using the features. The output of software releases is learning, not features.

This also re-frames the definition of “done”. We’re not done because we delivered the features. We’re only done when we’ve achieved the goal. Maybe we do that in one iteration. Maybe we need more throws of the dice to get there.

So this model of software development sees cross-functional teams working as one to achieve a goal, potentially making multiple passes at it, and working one goal at a time. The goal defines the team. “We’re the team who enables diners to find great Chinese food”. “We’re the team who puts guitar players in touch with drummers in their town.” “We’re the team who makes sure patients don’t forget when their repeat prescriptions need to be ordered.”

Maybe you need a programmer to make that happen. Maybe you need a web designer to make that happen. Maybe you need a database expert to make that happen. The team is whoever you need to achieve that goal in the real world.

Now I look at the current state of Agile, and I see so many teams munching their way through a list of features, and so few teams working together to solve real end users’ problems. Most teams don’t even meet real end users, and never see how the software gets used in the real world. Most teams don’t know what problem they’re setting out to solve. Most teams are releasing rotten tomatoes, and learning little from each release.

And driving all of this, most teams have a dedicated person who manages that backlog of features, tweaking it and “grooming” it every time the plan needs to change. This is no different to my experiences of updating detailed project plans in the 90s. It’s plan-driven development, plain and simple.

Want to get unstuck from working through a detailed long-term plan? Then ditch the backlog, get with your real customer, and start solving real problems together in the real world.