Benchmark method.
The fundamental shift replaces sequential planning with dynamic feedback loops. The Manager creates high-level tasks, the Executor takes one concrete action, then the Manager immediately reassesses.
The Manager–Executor loop.
A tight plan → execute → observe → reassess cycle, instead of plan-then-execute.
Sequential planning
Dynamic feedback loop
Manager creates a high-level task
Executor takes one concrete action
System observes environment changes
Manager reassesses and replans
Six technical improvements.
The innovations behind our top-tier performance on AndroidWorld.
Specialized Text Agent
Routes text-intensive tasks to a dedicated agent with Python shell access. Receives accessibility trees plus current text context and can atomically clear and replace content.
- Routes when Manager tags items as text tasks
- Receives accessibility trees and focused element context
- Atomic clear/replace for reliable text editing
Contextual Awareness
Enhanced through device date injection, 0.5-second screen stabilization waits, disabled pointer visualization, differential state tracking, and automatic app capability extraction.
- Device date injected into context
- 0.5s screen stabilization before reads
- Differential state: current vs. previous accessibility tree
- Automatic app capability extraction
Transparent Communication
Executors output three components, thought process, chosen action, and description, all injected into Manager context for full decision rationale.
Memory System
Guidance scattered throughout system prompts with repeated context in multiple sections for consistent availability. Injections into both system prompt and final user message.
Expanded Actions
Eight action primitives covering the full interaction surface of mobile devices, from simple taps to complex clipboard operations.
Prompt Engineering
Iterative refinement across system prompts through strategic distribution rather than concentration of instructions. Model-specific optimization patterns.
Eight action primitives.
Covering the full interaction surface of a mobile device.
Tight feedback loops, task-specific routing, rich state observability, and dynamic replanning outperform rigid plan-then-execute models for mobile UI automation.
Peer-reviewed research.
The mobilerun agent architecture is studied in published academic work.
Curiosity Driven Knowledge Retrieval for Mobile Agents
Co-authored by Droidrun's Shahir Ali and mobilerun co-founder Niels Schmidt and published at the ACM Web Conference 2026, the paper introduces uncertainty-driven retrieval of app knowledge, organized as AppCards, the concept behind the framework's App Instruction Cards, for mobile GUI agents. Evaluated on AndroidWorld using the open-source mobilerun (droidrun) framework for stable execution, curiosity-driven retrieval reaches an 88.8% success rate with GPT-5, reported in the paper as an improvement over the prior published best of 84.5%, with consistent gains in task success and robustness across models and difficulty levels.
See the results.
The full benchmark, 91.4% success across 116 AndroidWorld tasks.