Gyoza

The problem

A business assistant needs more than generated content. It must know which account it is acting for, which actions are permitted and whether an attempted interaction actually completed.

How I approached it

Build a configurable product around an independent fork of the Jev/Wheerli execution engine. Keep business interests, writing examples and action permissions in a versioned profile. Apply policy checks at execution time and separate attempted actions from confirmed outcomes.

What I built

• Business-profile studio for topics, exclusions, writing examples and per-action permissions
• Runtime policy checks tied to the confirmed profile revision and bound account
• Device-session workflows using the inherited execution engine, with observation history and publishing journals
• Operational work on navigation recovery and delivery confirmation for real device sessions

Where it stands

The product is under active development. Device sessions have been exercised, but navigation reliability and delivery confirmation still need work. Unconfirmed actions are kept distinct from successful delivery; the prototype is not presented as a mature autonomous service.

What it taught me

Reliable agent execution depends on state and evidence. An attempted action cannot be counted as completed, and a retry must account for the possibility that the first attempt succeeded without confirmation.