OpenAI has upgraded its Responses API, signaling a shift towards more capable AI agents, according to VentureBeat. The updates, announced on February 10, 2026, include Server-side Compaction and Hosted Shell, designed to enhance the capabilities of AI models. These improvements aim to address the limitations of earlier AI agents, which often struggled with context and long-term memory.
The upgrades come as the industry explores alternative memory architectures for AI agents. One such approach, "observational memory," developed by Mastra, promises to cut AI agent costs tenfold and outperform Retrieval-Augmented Generation (RAG) systems on long-context benchmarks, VentureBeat reported. RAG systems, which retrieve context dynamically, have proven insufficient for complex, long-running agentic AI workflows.
The evolution of AI agents has sparked discussions about their potential to reshape industries, including Global Business Services (GBS). While the promise of agentic AI, capable of goal-driven action, has been widely discussed, its actual deployment has lagged. VentureBeat Contributing Editor Taryn Plumb noted in a December 2025 post that the fundamentals needed for scaling agentic AI were still missing, drawing on input from Google Cloud and Replit.
The development of AI agents is also facing technical challenges. One user on Hacker News reported difficulties with tool calling when using GPT-OSS-120b, highlighting the need for proper implementation of inference engines. The user noted that the recent versions of llama.cpp handle tool calling well, but issues persist with additional quantization.
MIT Technology Review's new AI newsletter, "Making AI Work," is exploring the practical applications of AI across various sectors. The newsletter examines specific use cases and provides context on how companies and industries are employing AI tools.
Discussion
AI Experts & Community
Be the first to comment