Community · E2 · artifact verified
Drive an Android phone from the A11Y tree
An Android automation agent with a two-tier brain: Jev decides fast from the accessibility tree, a vision agent takes over only when the structural view is ambiguous, and ADB executes - cheap judge on the hot path, expensive one on the exceptions.
01 · Role in the system
What Jev does here
Phone automation usually burns a vision model on every step. This agent inverts the cost: the accessibility tree is already a structured description of the screen, so Jev answers typed questions over it - which element to tap, whether the screen state matches the goal - as the default decision maker. Only when the A11Y view is ambiguous does the visual agent take over, looking at pixels. Execution goes straight through ADB, and the split is measured: the cheap judge handles the routine steps, the expensive one handles the exceptions, and the Python package includes a probe script for checking the Jev path in isolation.
02 · Control boundary
Where Jev sits
Jev as the fast tier of a two-tier phone automation brain: typed decisions over the accessibility tree drive routine steps through ADB, and escalation to a vision agent is reserved for the ambiguous cases.
Code owns the loop, permissions, thresholds, validation, and side effects. Jev owns only the bounded judgments described above.
03 · Known limits
What this evidence does not prove
- A11Y coverage varies by app; apps with custom rendering escalate to vision more often.
- 4 stars; escalation rate between the two tiers is not published.
- Requires ADB debugging enabled on the device.
04 · Attribution
Public sources
This is a Community record: the project was published by a third-party community author.
- xinwang-nwpu ↗Community · github · public · checked 2026-10-06