Browsing by Author "Upaa Jr, Tersoo"

Advances in large language models have enabled the development of virtual agents that solve problems across practical, real-world applications. While these models have shown remarkable capability in optimizing for near-term progress, they struggle under the unique demands imposed by long-horizon settings. Few existing benchmarks emphasize partial observability, game-theoretic reasoning, and long-horizon planning simultaneously, leaving a gap between what evaluations measure and what real-world autonomy demands. Our work presents three contributions toward closing that gap. First, we introduce the PokéAgent Speedrunning Benchmark, a standardized evaluation framework set in Pokémon Emerald that requires agents to sustain coherent reasoning across thousands of interactions within a dynamic, partially observable environment. We further develop an expert agent harness that decomposes gameplay into a multi-agent orchestration system with purpose-built memory, planning, skills, subagents, and self- reflection capabilities. Tracing our expert harness through several stages of development, we identify key architectural decisions that most impact long-horizon performance. Finally, we introduce AutoEvolve, a reset-free agent-scaffolding algorithm that builds its own harness through mid-episode optimization. We evaluate five harness configurations over 24-hour evaluations using Gemini 3.1 Pro. Evolved harnesses consistently match or exceed expert efficiency through approximately 80% completion on the Pok´eAgent Speedrunning Benchmark and autonomously develop functional skills, tools, and subagents. Bootstrapped transfer accelerates early-game progress by as much as 20%, though frozen-harness runs degrade at higher difficulty, demonstrating the need for continuous self-adaptation to environmental dynamics.