Feature Flags: A Safety Net for AI Development

Foreword

Generative AI exploded into our collective consciousness with the launch of ChatGPT in 2022. Initially, I was sceptical of practical use cases—it was a neat parlour trick, but not much more. Early coding assistants' autocomplete suggestions were sometimes shockingly impressive, but overall they got in the way more often than they helped.

This changed quickly. By 2024, the capabilities of AI models had grown significantly, and the introduction of agents delivered a dramatic shift in what AI-assisted coding could actually accomplish.

Now, in early 2026, it's clear that AI-assisted engineering represents a tectonic shift in our industry—fundamentally altering the economics of software. Writing code just got a lot cheaper.

But building software involves a lot more than writing code. I spend a lot of time talking to engineering leaders, and one of the most prominent themes in those conversations is that AI has not dramatically reduced the overall cost of delivering features. It's just moved the bottleneck. Engineers are using AI to produce code faster, but those changes are now hung up in code review, testing, or deployment.

This is the big challenge facing engineering leaders in the AI era: how do we optimise the full value stream from ideation, through implementation, and on to real customer value?

In this book, we tackle that problem head-on.

Progressive Delivery for AI-Led Development

AI is having a huge impact on how engineers build software. But when viewed through the lens of the full software delivery lifecycle, the productivity boost from AI has been mostly limited to the implementation phase.

Think of the SDLC as a production line in an industrial bakery. The arrival of AI-assisted engineering is like upgrading the bread oven in the middle of the factory. You can bake dough faster—but unless you improve the rest of the line, you won't gain the improvements you might hope for from the system overall.

Solving these systemic optimisation problems is a common challenge in the world of physical manufacturing. For the last half-century, the standard toolkit has been Lean Manufacturing. In software, pioneering organisations have taken those Lean ideas and applied them to the SDLC—leading to a set of practices known as Progressive Delivery.

Progressive Delivery is built around two core principles: reducing lead time, and maximising learning. By optimising our SDLC to reduce lead time, we increase throughput and can deliver more features. By layering in techniques like A/B testing, we can be sure those features are actually valuable.

Central to all of this is the feature flag—a dynamic if/else statement that lets us change control flow at runtime, turning a feature on or off without deploying code. It's a deceptively simple tool with far-reaching consequences for how teams ship.

AI-Accelerated SDLC

"Toil," as defined by Google engineers in the Site Reliability Engineering book, is work that is manual, repetitive, automatable, and devoid of enduring value. In the context of software delivery, it's the rote tasks involved in moving a feature through your delivery process—starting a ticket, creating a feature flag, merging a PR, marking something as ready to test.

It's also exactly the kind of work that coding agents are well-suited to handle.

In this chapter, we walk through what it looks like to use AI to reduce friction and toil across the SDLC, using feature flag management as the illustrative use case. We follow a web experience team at a large bank building a "go paperless" widget, from adding the initial feature flag through to cleanup, and show how AI can reduce the overhead at each step.

The numbers tell a compelling story. Across all the steps involved, total toil drops from roughly six hours to around 23 minutes—a shift in the cost/benefit calculation of when to apply progressive delivery practices at all.

When putting a feature behind a flag becomes cheap by default, teams experiment more. They learn faster. And over time, that compounds into the kind of product velocity that was previously the exclusive domain of well-resourced big tech platforms.

AI-First Experimentation

AI has now evolved to the point where someone without programming experience can bring working software into existence just by describing it. You prompt a coding agent, watch it build an app from scratch, then refine functionality by continuing to chat. This is "vibe coding"—placing your full trust in the agent without reviewing the code it creates.

The first time you try it, vibe coding can be a jaw-dropping experience. However, software created this way has real limitations. Coding agents aren't yet capable of making all the design decisions required to build secure, reliable, maintainable systems. And vibe coding only really works in greenfields—when an agent is creating an app from scratch, or making changes to a small codebase it can understand in its entirety.

Despite these limitations, it shouldn't be dismissed as something only suitable for toy apps.

One useful mental model is blazing a trail, then laying the rail. Product managers explore rapidly with vibe-coded experiments, bushwhacking through the undergrowth to find the best route. Then engineers follow, replacing that rough trail with a structurally sound railroad—one that can carry a heavy load for years.

Critically, this only works safely if progressive delivery infrastructure is already in place. Feature flags are what allow throw-away experiments to live, briefly, in production systems, without putting those systems at risk.

Wrapping up

AI dramatically increases our capacity to produce code. What it doesn't automatically do is improve an organisation's ability to deliver value.

Bottlenecks that were once hidden by the expense of writing code are now exposed: review queues, testing delays, release friction.

Feature flags, and the broader set of Progressive Delivery practices they enable, are a critical part of the solution: decoupling deployment from release, shrinking batch sizes, and enabling continuous experimentation.

In an AI-accelerated world, where changes are cheaper and more frequent, these practices are no longer aspirational. They're table stakes.