AGI + Krea: Designing control for AI phone agents and image tools

Problem

AGI's Android agent operates your phone for you. Without native hardware, people had to work through custom onboarding and permissions before they could even turn it on.

What I designed

• Onboarding prototypes that made manual activation as easy as possible, drawing on the patterns in Nikita Bier's Explode app
• A state machine that lets people steer the agent when it gets stuck or hallucinates

What I learned

• Consumers expect responses under 200 ms and 99.999% reliability. In internal testing the agent fell short of that, so I believe agent phones are still years from the mainstream
• A prompt adds far more noise between intent and action than a tap does

Status pills for running, paused, retry and complete, each with the next action in reach.

Context

Midjourney ran on Discord slash commands. Krea's bet was the canvas: generate images, then organize and refine them visually on the web. ControlNet then let people guide generation with edges, depth maps and pose skeletons, and the whole field raced to support it.

What I designed

• Information architectures for working with images and prompts on the canvas
• Entry points for prompt + image + mask
• Loading and error states

Boom times in San Francisco - The Information (Press)