OpenMobile-2: Open-Source Smartphone Agents Trained Across Many Apps

What happened

Researchers behind OpenMobile-2 report that open-source smartphone agents can perform strongly on real Android devices, scoring highest among open agents in the authors' testing, according to The Decoder's coverage. The headline claim that the system rivals proprietary models is the outlet's framing; the reporting attributes the performance results to the authors.

The team's central argument is that smartphone agents improve by training across many different applications rather than by processing ever-larger volumes of data. To test this, they built an open training environment containing more than 110 applications, including 35 replicas of popular apps, giving agents a broad range of interfaces to learn from.

The reporting does not give a head-to-head benchmark figure against proprietary systems in the summary material provided, so the size of any gap with commercial models remains unclear from the coverage alone.

Why it matters

Smartphone agents act on a user's behalf by tapping, typing and navigating across apps. What OpenMobile-2 emphasises is that the quality of that navigation depends heavily on exposure to varied, real-world interfaces. For organisations, this shifts the question from how much data an assistant has ingested to how many distinct digital journeys it has been trained to complete reliably.

For leaders in digital transformation, the development points to a near-term change in how services are reached. If capable open agents can operate across third-party apps, the front door to a service may increasingly be an agent acting for the customer, rather than a brand's own app or website. That has implications for how services are designed, tested and governed.

By the numbers

• More than 110 applications included in the open training environment.
• 35 replicas of popular applications built into that environment.

The Renascence take

Most coverage of this kind will focus on the leaderboard. The more useful signal is the training philosophy: breadth of interface beats depth of data. An agent that has seen 35 variations of popular apps is learning the messiness of real digital products, the inconsistent labels, hidden menus and unexpected pop-ups that customers themselves struggle with.

The next competitive edge in digital service may not come from owning the best AI model, but from making your interfaces legible to the agents acting for your customers. Brands that treat their app as a closed, pixel-perfect experience will find agents stumbling at the seams. Operators should audit their own journeys the way an agent would: looking for ambiguous labels, fragile flows and steps that only make sense to someone who already knows the product. Designing for agent-readability and for clear recovery when something fails is becoming a customer experience discipline in its own right.