In an impressive feat, Japanese startup Sakana AI’s <a href="https://sakana.ai/ahc058/">coding agent ALE-Agent</a> recently secured first place in the AtCoder Heuristic Contest (AHC058), a complex coding competition that involves complicated optimization problems — and a more difficult and perhaps telling challenge than benchmarks like HumanEval, which mostly test the ability to write isolated functions, and which many AI models and agents now regularly pass with ease (&quot;benchmark saturation&quot;). Sakana&#x27;s accomplishment with ALE-Agent hints at a shift toward agents capable of autonomously optimizing themselves to navigate and perform well in complex, dynamic systems such as enterprise software stacks, workflows, and operational environments. In four hours, the agent used inference-time scaling to generate, test, and iterate over hundreds of solutions, solving a problem that typically requires deep intuition and time-consuming trial and error from human experts. It outperformed over 800 human participants, including top-tier competitive programmers.<h2>How ALE-Agent works</h2>The challenge in AHC058 was a classic <a href="https://en.wikipedia.org/wiki/Combinatorial_optimization">combinatorial optimization</a> problem. Participants were tasked with managing a set of machines with hierarchical relationships, such as machines that produce apples, and other machines that build those apple-producing machines. The goal was to maximize output over a fixed number of turns.In the enterprise world, this workflow usually follows a strict pattern: a domain expert works with a client to define an &quot;objective function&quot; (aka the Scorer), and then engineers build a software system to optimize it. These problems are notoriously difficult because they cannot be solved in a single stage. They require exploration, strategy, and the ability to pivot when a plan isn&#x27;t working.Human experts typically approach this using a two-stage strategy. First, they use a &quot;Greedy&quot; method (a lightweight solver that makes the best immediate choice at each step) to generate a decent baseline solution. Then, they apply &quot;<a href="https://en.wikipedia.org/wiki/Simulated_annealing">simulated annealing</a>,&quot; a technique that takes the existing plan and makes tiny, random adjustments to see if the score improves. However, this standard approach is rigid. If the initial Greedy plan heads in the wrong direction, simulated annealing can rarely fix it because it only looks for local improvements in a faulty area of the solution space.ALE-Agent’s innovation was transforming this static initialization tool into a dynamic reconstruction engine. Instead of relying on immediate value, the agent independently derived a concept it called &quot;Virtual Power.&quot; It assigned values to components that were not yet operational, treating them as if they already possessed value. By valuing potential future assets rather than just current ones, the agent capitalized on the &quot;compound interest effect,&quot; a concept it explicitly identified in its <a href="https://sakanaai.github.io/fishylene-ahc058/">internal logs</a>. Basically, it could look a few steps ahead and reason about the future instead of looking at the immediate feedback it was receiving from its environment.Crucially, the agent needed to maintain this strategy over a four-hour window without losing focus, a common failure mode known as “context drift.” In comments provided to VentureBeat, the Sakana AI team explained that the agent generates textual &quot;insights&quot; by reflecting on each trial. It gathers this knowledge to prevent cycling back to previously failed strategies and creates a working memory that allows it to look a few steps ahead rather than just reacting to immediate feedback.Furthermore, the agent integrated Greedy methods directly into the simulated annealing phase to avoid getting stuck in local optima, using high-speed reconstruction to delete and rebuild large sections of the solution on the fly.<h2>From coding to enterprise optimization</h2>This breakthrough fits directly into existing enterprise workflows where a scoring function is already available. Currently, companies rely on scarce engineering talent to write optimization algorithms. ALE-Agent demonstrates a future where humans define the &quot;Scorer&quot; (i.e., the business logic and goals) and the agent handles the technical implementation.This shifts the operational bottleneck from engineering capacity to metric clarity. If an enterprise can measure a goal, the agent can optimize it. This has direct applications in logistics, such as vehicle routing, as well as server load balancing and resource allocation.According to the Sakana AI team, this could democratize optimization. &quot;It enables a future where non-technical clients can interact directly with the agent, tweaking business constraints in real-time until they get the output they desire,&quot; they said.The Sakana AI team told VentureBeat that ALE-Agent is currently proprietary and not available for public use, and the company is currently focused on internal development and proof-of-concept collaborations with enterprises.At the same time, the team is already looking ahead to &quot;self-rewriting&quot; agents. These future agents could define their own scorers, making them feasible for ill-defined problems where human experts struggle to formulate clear initial metrics.<h2>The cost of intelligence</h2>Running ALE-Agent was not cheap. The four-hour operation incurred approximately $1,300 in compute costs involving over 4,000 reasoning calls to models like <a href="https://venturebeat.com/ai/gpt-5-2-first-impressions-a-powerful-update-especially-for-business-tasks">GPT-5.2</a> and <a href="https://venturebeat.com/ai/google-unveils-gemini-3-claiming-the-lead-in-math-science-multimodal-and">Gemini 3 Pro</a>. While this price point might seem high for a single coding task, the return on investment for optimization problems is often asymmetric. In a resource-management setting, a one-time cost of a few thousand dollars can result in millions of dollars in annual efficiency savings.However, enterprises expecting costs to simply drop might be missing the strategic picture. While the cost of tokens is falling, total spend may actually rise as companies compete for better answers, a concept known as the <a href="https://en.wikipedia.org/wiki/Jevons_paradox">Jevons paradox</a>.&quot;While smarter algorithms will drive efficiency, the primary value of AI is its ability to explore vast solution spaces,&quot; the Sakana AI team said. &quot;As inference costs fall, rather than simply banking the savings, enterprises will likely choose to leverage that affordability to conduct even deeper, broader searches to find superior solutions.&quot;The experiment highlights the immense value still to be unlocked through inference-time scaling techniques. As AI systems gain the ability to handle complex reasoning tasks across longer contexts, building better scaffolding and allocating larger budgets for &quot;thinking time&quot; allows agents to rival top human experts.

حققت شركة ساكانا إيه آي اليابانية الناشئة المركز الأول في مسابقة أتكودر هيوريستيك (AHC058) مع وكيل البرمجة الخاص بها، ALE-Agent، مما يظهر قدرته على حل مشكلات التحسين المعقدة بشكل أكثر فعالية من المعايير التقليدية مثل HumanEval.

La startup japonesa Sakana AI logró el primer lugar en el Concurso Heurístico de AtCoder (AHC058) con su agente de codificación, ALE-Agent, demostrando su capacidad para resolver problemas de optimización complejos de manera más efectiva que los benchmarks tradicionales como HumanEval.

La startup japonaise Sakana AI a remporté la première place au concours heuristique AtCoder (AHC058) avec son agent de codage, ALE-Agent, démontrant sa capacité à résoudre des problèmes d'optimisation complexes plus efficacement que les benchmarks traditionnels comme HumanEval.

Japanese startup Sakana AI's coding agent, ALE-Agent, achieved first place in the AtCoder Heuristic Contest (AHC058), showcasing its ability to solve complex optimization problems more effectively than traditional benchmarks like HumanEval.

Why Sakana AI’s big win is a big deal for the future of enterprise agents

One More Thing in AI – Your Shortcut to AI Mastery

Why Sakana AI’s big win is a big deal for the future of enterprise agents

Was this article worth reading? Share it

One More Thing in AI

Chattermate

Kansei

Synthx

KushoAI

CodeRank

Ready to build your own newsroom?