How to build a Self-improving Product

🧠 Is this the future of product development? Real world examples with templates and guides to follow.

🔒 The Knowledge Series breaks down emerging AI technologies with practical playbooks designed specifically for product teams. Get 100+ guides and practical tutorials covering everything from Claude Code and MCP to agentic workflows, vibe coding, and more.

---

There’s a new concept that is starting to gain traction in tech circles and that is the idea of a self-improving product.

YC partner and ex-Monzo CEO Tom Blomfield gave a fascinating talk on the concept and earlier this month and Amplitude recently released a new feature in a closed beta called Wave which it explicitly markets as a tool that “enables self improving products”.

Here’s Tom Blomfield explaining the concept in a bit more detail:

The idea is pretty straightforward: what if you could deploy swarms of agents who constantly monitor your product’s performance, spot gaps, bugs or UX issues and proactively suggest - and then implement - those fixes, with minimal human intervention?

That is the essence of a self improving product.

And it hints at a future where agents will not only disrupt the work of “doing” i.e. development / design, but also the decision making layer above that that many of us assumed might be the last piece of the development process that really required human judgement.

But how realistic is all of this? And what impact could it have on the future of product development? Amplitude says their ship rate of AI agent product feature suggestions is still low at around 11%, but it’s not unreasonable to assume that this will grow over time.

In this Knowledge Series, we’re going to unpack some of the core themes and trends that are currently emerging around this topic and lay out a step by step guide on how to build your own self improving product system.

---

The Knowledge Series

What a self-improving product system looks like - a snapshot of the core parts

Here’s a snapshot of what a self improving product loop might look like:

This type of system has five core architectural parts, each of which performs a dedicated function:

1. Product map - the foundation everything runs against, defining your product’s core areas, user journeys, and success metrics

2. Sense - pulls data from multiple sources (Mixpanel, Amplitude, FullStory, Zendesk) and returns a list of potential problems and solutions

3. Decide - a quality judge scores and filters opportunities, scans the codebase via GitHub, and writes the tech spec

4. Act - takes the spec and routes it to the right executor: for example, Claude Code for autonomous fixes, Linear or Jira for human-picked-up work

5. Learn - measures the impact of what shipped and updates system memory so every subsequent run is better informed than the last

With tools like Claude Code, we can turn this into a loop that we can run on demand.

A closer look at this in more detail and how to build your own self-improving product system

Coming up:

• A closer look at what Amplitude built and the core concepts

• How to build your own in-house version of this system; a detailed step by step breakdown and blueprint

• The core pieces of the loop explained with prompts, agent templates and Claude Code folder structures you can use