AlgoEvo: Self-Evolving Agentic Search for Automated Algorithm Discovery

Automated Algorithm Discovery (AAD) aims to generate high-quality optimization heuristics without requiring manual engineering. While Large Language Models (LLMs) have made strides in this field, existing frameworks often rely on rigid, pre-defined pipelines that treat the LLM as a passive code sampler. This leads to "amnesic" search processes that discard valuable execution feedback and struggle to transfer knowledge between different types of tasks. AlgoEvo addresses these limitations by introducing an agentic framework that transforms algorithm discovery into an interactive, knowledge-accumulating process, allowing for more efficient and adaptable design.

An Interactive Agentic Loop

Instead of following a fixed, static schedule, AlgoEvo employs an autonomous agent that actively manages the discovery process. This agent dynamically decides when to inspect code, perform diagnostic analyses, edit algorithmic components, or evaluate new candidates based on real-time feedback. By interleaving these actions, the agent can address bottlenecks as they arise and perform structural refactoring, moving away from simple trial-and-error routines toward a more deliberate, scientific approach to algorithm design. The ai agents story also surfaces in Stanford Researchers Develop TRACE to Fix..., adding another angle.

Decoupling Design Skills

A major challenge in automated design is the variety of paradigms, such as single-objective optimization, multi-objective trade-offs, and multi-component systems. Traditional methods often require separate, fragmented codebases for each. AlgoEvo solves this through a "design skill hub," which decouples paradigm-specific knowledge from the core discovery engine. This hub acts as a pluggable interface: by simply supplying the appropriate skill, the same agentic engine can seamlessly handle diverse design tasks without requiring structural modifications to the underlying search loop.

Hierarchical Experience and Self-Evolution

To ensure that the system improves over time, AlgoEvo uses a hierarchical experience mechanism. As the agent explores, it distills its actions and performance outcomes into "experience cards." These cards are organized into a task-level tree, which the agent uses to guide future decisions through a selection process that balances exploration and exploitation. Beyond individual tasks, the framework performs cross-task consolidation: when effective design patterns are identified across multiple problems, they are abstracted into the skill hub. This allows the system to accumulate reusable knowledge, ensuring that later tasks benefit from the expertise gained in previous ones. The ai agents story also surfaces in Andrew Ng Launches OpenWorker to Deliver..., adding another angle.

Performance and Efficiency

Experiments across six representative benchmark tasks—ranging from single-objective heuristics to complex multi-component systems—demonstrate that AlgoEvo is highly effective. The framework consistently matches or surpasses specialized state-of-the-art methods while requiring significantly fewer environment evaluations and lower token consumption. These results highlight the framework's ability to achieve superior performance through flexible skill activation and the continuous accumulation of algorithmic experience. The ai agents story also surfaces in Claude autonomously improved models across 10..., adding another angle. as detailed in the full paper on Arxiv