Agents

The tempting version of an Android agent is simple: give it access to a phone, let a model inspect the screen, and ask it to do things. The problem is that “do things” is not a permission boundary. I’m building Argo, a control plane for Android actions, around a narrower idea: the phone should execute signed, typed jobs. A job says what is allowed to happen. The device verifies it and performs that bounded action. It does not need a language model living on the phone, deciding freely what should happen next. ...

Argo is an Android automation project built around signed, typed jobs rather than an agent with unlimited discretion. A host prepares a narrow action request; the companion verifies it and returns a structured result. Status: Active development Stack: Android, Flutter, Python, Hermes tooling

Most AI tools are about thinking faster. You describe a problem, get a response, revise, repeat. The loop is still you doing the work — just with a smarter clipboard. Agents are different. An agent can receive a goal, figure out what needs to happen, take actions, check results, and adjust. The loop runs without you in it. That shift is smaller than it sounds in demos and larger than it sounds in practice. ...