Reach new heights of laziness by burning tokens to let an AI tap your smartphone screen for you

Opposable, a new AI tool for controlling smartphone apps, took 50 seconds and 17 tool calls to switch to dark mode and set a five-minute timer; another demonstrated lookup cost $0.44 and took nearly two minutes. In tests on eight untuned tasks, the model completed 4 and 3 tasks across two runs, rising to 6 in each run when guided one subgoal at a time. The Android app requires sideloading and screen-access permissions, and its developers warn that anyone holding the access token can operate the phone.

Analysis

The investment signal is not this product’s traction; it is the friction between general-purpose agents and mobile operating systems. A brittle, permission-heavy third-party agent raises the value of platform owners’ ability to control device APIs, identity, and user consent. That is directionally supportive of AAPL’s iOS gatekeeping and could favor GOOG if Android can offer useful agent capabilities through trusted native interfaces rather than broad accessibility access. This is a conditional platform advantage, not evidence of near-term revenue or margin uplift for either company.

The larger adoption constraint is reliability and trust, not the availability of another model. When an agent can read screens and act as the user, a single mistaken action or leaked credential can outweigh modest time savings. That raises the bar for consumer deployment and creates a potential opening for enterprise-managed workflows and accessibility use cases, where permissions and task value are clearer. Conversely, the showcased limitations are from a small, narrow test and should not be extrapolated to the capability of multimodal agents generally.

Near term (days to weeks), no material fundamental read-through to AAPL or GOOG; treat the story as sentiment noise. Over 1–3 months, watch native agent launches, permission changes, and evidence of repeat use rather than demos. Over 6–18 months, the competitive prize is control of the trusted action layer: OS-native execution could retain user engagement and interface leverage, while independent agents remain vulnerable to platform restrictions. The contrarian risk is that a step-change in reliability or standard APIs rapidly removes today’s friction.

Key Decisions for Investors

• No trade on this launch alone. It does not establish adoption, monetization, or an earnings impact for AAPL or GOOG.
• Watch for a relative-value opportunity only if evidence emerges that native agent capabilities improve engagement or paid-service usage: compare AAPL and GOOG after verifying product availability, usage, and management commentary rather than assuming either platform captures value.
• Track platform policy and security incidents as catalysts. A major agent-related privacy breach or tighter restrictions on accessibility/API access would reinforce the trusted-platform thesis; open, reliable cross-platform action APIs would weaken it.
• Falsify the current cautious adoption view with independently measured repeat usage, materially higher task-completion reliability, and reduced user oversight. Until then, treat consumer agent demos as product experimentation, not a commercial inflection.