Building Miyagi — What It Took to Make AI Feel Useful in a Real Product

October 11, 2025

When we started building Miyagi at Amply, the goal wasn't to showcase AI. It was to remove friction from work people already do inside Webflow.

Most AI products fail not because the tech is weak, but because the product thinking stops at "look what it can do." We wanted Miyagi to feel less like a demo and more like a teammate that knows its limits.

Agents should have jobs, not superpowers

Instead of a single do-everything assistant, Miyagi is built around clear responsibilities — design help, content updates, CMS work, and code-related tasks, all within Webflow's context.

This reduced unpredictable behavior and made outcomes easier to trust.

Users don't want intelligence. They want reliability.

Context beats intelligence

Raw AI capability mattered less than where the agent operated.

Because Miyagi works inside Webflow:

  • it understands real structure
  • actions are visible
  • feedback is immediate

The biggest unlock wasn't smarter prompts. It was tighter context.

Automation should feel optional

We designed Miyagi to assist, not override. It proposes changes. Users decide what ships.

That choice slows things slightly — and that's good.

Speed without control creates anxiety. Predictability creates confidence.

The real work lives in the edges

Partial failures, undo flows, unclear intent — these weren't AI problems. They were product problems.

Treating Miyagi like a product first, and an AI system second, is what made it usable.