Agenda

The shift: from building to validating

Why the traditional PRD-first workflow is breaking down, and how teams are now putting prototypes in front of real users before committing engineering time.

AI-powered products need constant iteration

If your product includes AI (recommendations, copilots, assistants, generated content), you’re no longer shipping fixed features—you’re shipping systems that need tuning:

• Prompt changes
• Model switches
• Output variations
• Behaviour tweaks

We’ll show how teams are using feature flags to control and iterate on these in production, safely and instantly.

The new bottleneck: learning speed

You can build anything quickly, but most ideas still won’t land. The advantage has shifted to teams that can:

• Test ideas with real users immediately
• Measure impact fast
• Double down or kill features without friction

Feature flags as your iteration layer

How teams are using flags as the infrastructure for this new model:

• Release prototypes to targeted cohorts
• Compare different versions (including AI behaviours)
• Roll back instantly if something doesn’t work
• Continuously refine features after launch—not just before

Making this work across engineering teams

How to operationalise this without adding friction:

Standardise flags so every feature is shippable by default

• Embed them into AI-assisted workflows (so everything comes flagged out of the box—even for new engineers)
• Integrate into pipelines (e.g. MCP + GitHub Actions) to keep releases fast and consistent
• Add lightweight controls (like change requests) that scale with high-velocity shipping

The Shift: What Matters Now

1. The experimentation opportunity is massive now

Not everything belongs in production experiments — if you're building payment reconciliation at scale, that's pure logic, you need rigorous engineering and PRDs, and that's not changing. But there's this entire other side that most teams are barely touching. Think Booking.com: their whole UX is a science of hypothesizing and discovering what makes people move through the funnel. You can't logically deduce what makes a human click — you have to test it, learn, iterate. They've built an empire on that. The difference now is you don't even need to prioritize anymore. Spinning up an experiment is nearly free.

2. Agents need this even more

If a traditional UI is already hard to reason about from first principles, agents multiply that by orders of magnitude. The problem space is basically infinite — you're dealing with natural language, unpredictable user intent, and behavior you simply cannot predict upfront. Experimentation isn't a nice-to-have for agents, it's the only effective way to improve them. And this applies everywhere, even inside regulated companies. Adyen isn't going to experiment with their payment code, but if they have an internal risk analysis assistant with a human in the loop? You absolutely can run experiments there — analysts give feedback on correctness, you measure whether it's saving them time, and you iterate. That's the opportunity.

Live Demo

Q&A