Building a Pocket AI Phone — and Learning How Hard Reliable Agents Really Are

I’m working on a small AI phone called Meydo AI.

The idea came from something pretty simple: I use AI every day, but on a phone it still often feels like opening another app, asking a question, and then doing the actual work myself.

We’re trying to make that interaction more direct. The device is small enough to carry around, and the agent can use voice, vision and supported tools to help complete tasks instead of only returning an answer.

The hardest problem for us right now is reliability. Getting an agent to complete a perfect demo is one thing. Getting it to behave properly when permissions change, a login expires, an app changes, or the user interrupts halfway through is much harder.

That’s probably the part I’m spending the most time thinking about right now.

Curious how other people here are handling reliability and confirmation for agents that can actually take actions.