For the past two years, every AI Agent has been "squatting" in someone else's operating system—making do with screenshots, simulated clicks, and API calls. AOHP, an overlooked research project, makes this explicit: Android must be rebuilt for AI. We see this not as a product launch but as an exam paper the academic world has handed to Google.
What this is
AOHP (Android Open Harness Project) is a research project built on the open-source version of Android 16 (AOSP). Two terms: Agent refers to an AI program that can autonomously plan multi-step tasks; OS (operating system) is the lowest-level management software on a device.Traditional Android is designed around the "human-app" model: one app at a time, eyes watching, fingers tapping, and cross-app interaction via share menus. AI Agents need to operate multiple apps simultaneously, remember user preferences, avoid re-recognizing interfaces from scratch every time, and—critically—never let AI read raw passwords.AOHP does three things at the system layer: lets AI assemble capabilities across apps (service aggregation); uses a "virtual display" to let AI run multiple interfaces in parallel in the background, several times faster than screenshot simulation (efficient interfaces); and lets AI invoke sensitive data without seeing the original—data lives in a protected "vault" (security isolation). The code is open-sourced on GitHub.
Industry view
Supporters see this as foundational infrastructure for Agent deployment. Today's Agent products all patch things at the application layer—calling APIs, relying on RPA (software that simulates human operation) to click buttons—which is inefficient and insecure. Without OS-level change, Agents will remain toys.But cooler heads push back. Critics flag three points:First, the power to modify the OS is highly concentrated in Google; whether the academic proposal gets adopted is an open question. Second, the paper's authors acknowledge this is only an Android-direction modification—if Agents eventually run on AR glasses, car systems, or robots, its reference value is limited. Third, security promises stay at the paper level; zero-day vulnerabilities (security flaws the vendor hasn't yet discovered) and side-channel attacks (eavesdropping via indirect signals like power consumption) have no answers.What we find worth watching: Apple, Huawei, and Xiaomi—the vendors holding the hardware + OS + AI trinity—as of now, none has publicly responded.
Impact on regular people
For enterprise IT: permission boundaries designed around "people" no longer apply; permission management may need to be rebuilt.For individual professionals: walls between apps will loosen; within three years we may see experiences like "one sentence telling AI to book flights across apps and write expense reports"—provided OS vendors are willing to open up.For consumer markets: no visible impact yet; this remains a matter for academia and developers.
