Sakana AI announced that its AI agent, ALE-Agent, achieved 1st place in the AtCoder Heuristic Contest (AHC) 058, outperforming 804 human participants. This marks the first known instance of an AI agent winning a major real-time optimization programming contest. For context, an OpenAI agent previously secured 2nd place in the AHC world tournament last August.
Contest Challenge and ALE-Agent’s Approach
AHC058, held on December 14, 2025, presented a 4-hour production planning optimization problem. Participants needed to plan hierarchical machine production efficiently, determining which types and levels of machines to upgrade and in what sequence. While problem setters anticipated a standard approach combining constructive heuristics with simulated annealing, ALE-Agent independently discovered a more effective strategy.
The agent implemented a “virtual power” heuristic, along with diverse neighborhood search operations within simulated annealing. These techniques enabled it to escape local optima more effectively than human competitors. ALE-Agent ran parallel code generation, analyzed results in real time, and iteratively refined its algorithms—an approach leveraging inference-time scaling across multiple frontier AI models, including GPT-5.2 and Gemini 3 Pro Preview. The total cost of computation and infrastructure was approximately $1,300.
Performance and Analysis
ALE-Agent began submitting solutions two hours into the contest and quickly reached 1st place, maintaining the lead until the contest ended. Analysis of its approach revealed unique Greedy methods, large-scale neighborhood operations, and high-speed simulation optimizations that would have been prohibitively time-consuming for human participants. Logs indicate the agent repeatedly refined its strategy, integrating insights from prior trial-and-error cycles into subsequent algorithmic decisions.
Experts praised ALE-Agent’s performance. Hiroomi Nochide, AHC058 problem author, noted that the agent’s combination of trial-and-error and LLM reasoning provided advantages beyond typical human approaches. Yoichi Iwata, AHC overseer, highlighted ALE-Agent’s ability to restructure large portions of the investment plan mid-competition, surpassing human-expected strategies.
Significance and Future Directions
The victory demonstrates that AI agents can now match or exceed human experts in complex tasks requiring extended reasoning and algorithmic innovation. Sakana AI emphasizes that ALE-Agent is designed as a partner to human problem-solving, extending exploration and discovery rather than replacing human expertise.
Looking forward, Sakana AI aims to enhance ALE-Agent’s stability for longer-duration tasks and refine its balance between trial-and-error and human-like reasoning. The company also plans to expand its practical applications of AI agents in real-world optimization problems.
