This paper investigates the use of AI agents in professional software development in 2025. Experienced developers value AI agents as a productivity boost but employ strategies to control agent behavior using their expertise to maintain fundamental software quality attributes. They generally feel positive about incorporating agents, confident in their ability to complement agent limitations. The findings highlight the value of software development best practices for effective agent use, suggest suitable tasks for agents, and point to future opportunities for improved agentic interfaces and guidelines.

This study uses large language models (LLMs) to investigate long-range structure in English texts. It finds that conditional entropy (code length) continues to decrease with context length up to approximately 10^4 characters, implying dependencies across these distances. Different learning dynamics for long- and short-range contexts suggest gradual learning of long-range structure. The results constrain efforts to build statistical physics models of LLMs or language itself.

Meta has acquired Manus, a Chinese-founded, Singapore-based AI startup specializing in AI agents, for over $2 billion. The acquisition aims to accelerate Meta's integration of AI and automation capabilities across its products. Manus develops general-purpose AI agents capable of complex tasks like data analysis and coding, and will continue its subscription service post-acquisition. This deal is notable as one of the first acquisitions of a Chinese startup by a major US tech company.