AI Agent Wins First Place in Optimization Programming Contest

On December 14th, an AI agent called ALE-Agent, built by Tokyo-based Sakana AI, entered AtCoder Heuristic Contest 058 under the account name fishylene and simply won. Not top ten, not a strong showing — first place, ahead of 804 competitors, in a four-hour contest that Sakana says is the first time any AI has taken outright victory in a live optimization programming contest.

The problem itself sounds almost whimsical: a hierarchy of machines that produce apples, and machines that build those machines, where competitors had to figure out which layer to invest in and when. Strip away the fruit metaphor and you've got a stand-in for supply chains, food webs, or any staged production process — exactly the kind of real-world optimization AtCoder contests are designed to mimic. ALE-Agent started submitting two hours into the contest and immediately grabbed provisional first. It traded the lead back and forth with runner-up yosupo through the middle stretch before pulling ahead for good around the two-and-a-half-hour mark.

What's notable isn't just that ALE-Agent won, but how. The contest's own problem-setter, Yoshiomi Ushiro of ALGO ARTIS, had expected a two-stage approach: search broadly for a rough investment plan with a lightweight method, then refine it. ALE-Agent and yosupo both instead leaned on local search with a twist — a