what's new in ai agents & automation

ACM

uses llm agents with runtime feedback to handle code migrations across large codebases, catching subtle errors that static analysis misses. practical for teams managing library upgrades and systematic refactoring.

Bingzheng Li, Xiaokang Yin, Yao Zhang +2

an agent that automatically repairs kernel bugs using coverage guidance and context awareness to navigate complex codebases. demonstrates how agents can handle high-stakes, low-tolerance-for-error domains.

Kishan Athrey, Ramin Pishehvar, Brian Riordan +1

proposes methods to automatically compose multi-agent systems by recommending which agents should handle specific user intents. helps practitioners build flexible, reusable agent families without manual orchestration.

Cheonsu Jeong, Younggun Shin

addresses reliability challenges in llm-based agents by enabling them to detect and recover from failures without human intervention. critical for production deployments where agents must operate robustly in unpredictable conditions.

Zerun Ma, Guoqiang Wang, Xinchen Xie +6

uses agents to automatically explore and optimize llm fine-tuning configurations through tree-based search. reduces manual hyperparameter tuning effort for practitioners customizing models for specific tasks.

Seamus Brady

provides a persistent runtime for long-lived agents with auditability, memory management, and safety constraints built in. essential for practitioners deploying agents in regulated or high-accountability environments.

Nature

a system that autonomously navigates the entire research lifecycle from conception to publication, automating what has traditionally required human researchers. practitioners building ai systems can learn from how this work structures long-horizon task decomposition and iterative refinement at scale.