AgentProg: Empowering Long-Horizon GUI Agents with Program-Guided Context Management

AgentProg introduces a novel program-guided context management system for long-horizon GUI agents that addresses the critical bottleneck of expanding interaction history overhead. By reframing interaction history as structured programs with variables and control flow, the approach preserves semantic information while reducing context requirements, achieving state-of-the-art performance on AndroidWorld benchmarks while maintaining robustness on extended tasks.

AgentProg represents a meaningful advancement in autonomous agent architecture, tackling a fundamental efficiency problem that has limited practical deployment of long-horizon task automation systems. Mobile GUI agents face exponential growth in context requirements as task complexity increases, forcing difficult tradeoffs between model performance and computational cost. Traditional compression techniques sacrifice semantic fidelity, leading to performance degradation. The program-guided approach reframes this challenge by imposing structural discipline on interaction histories, enabling intelligent information retention based on program semantics rather than arbitrary compression rules.

This work builds on growing recognition that agent systems require fundamentally different architectures than language models alone. The integration of belief state mechanisms from MDP frameworks demonstrates sophisticated thinking about partial observability—a critical real-world constraint where agents operate with incomplete environmental information. The extensibility to unexpected environmental changes addresses practical deployment concerns that theoretical benchmarks often ignore.

The research impacts multiple stakeholder groups differently. For AI researchers, AgentProg provides a replicable framework for context management that may generalize beyond GUI automation to other sequential decision-making domains. Developers building autonomous systems gain immediate tooling through open-source release, reducing barriers to implementation. The robust performance on long-horizon tasks has practical value for mobile automation, customer service bots, and accessibility tools.

Future development likely explores whether program-guided context management applies to other domains—code generation, robotic manipulation, or multi-agent coordination. Subsequent work should examine how semantic preservation scales with task complexity and whether the approach requires task-specific engineering or generalizes across domains.

• →AgentProg uses program structure to intelligently manage interaction history, solving the context overhead bottleneck in long-horizon GUI agents
• →The approach integrates belief state mechanisms to handle partial observability and environmental changes beyond standard baseline capabilities
• →State-of-the-art results on AndroidWorld benchmarks demonstrate maintained performance on extended tasks where baseline methods catastrophically degrade
• →Open-source release enables rapid adoption by researchers and developers building autonomous mobile and automation systems
• →Program-guided context management may generalize to sequential decision-making domains beyond GUI automation