Daily AI Updates: November 26, 2025

Main Content

• SGLang dLLM Framework Release: Launches in collaboration with Ant Group, NVLabs/Fast‑dLLM, and UIUC. Adds Block diffusion, KV cache management, and TP/EP parallelism. Integrates Ant Group’s new LLaDA2.0‑mini and LLaDA2.0‑flash models, with CUDA graphs and dynamic batching slated for next week.

• SGLang FP8 Training & Inference Optimization: Implements end‑to‑end FP8 for reinforcement learning, unifying FP8 to eliminate quantization mismatch. Reports significantly lower TIS‑clipfrac and improved stability on 30B and 235B MoE architectures.

• LangChain Deep Agents Update: Ships a general agent framework supporting persistent execution, streaming, and human‑in‑the‑loop cycles—suited for context engineering and deep agent construction.

• JetBrains AI Agent Tools Extension: Adds a custom ExecuteShellCommandTool for Kotlin coding agents to run code, observe failures, and iterate based on real feedback.

• GitHub Copilot Chat Cookbook: Releases prompt strategies and best practices to improve AI pair‑programming workflows.

• Anthropic AI Agent Harness Update: Inspired by human engineer workflows, optimizes long‑running agent execution across multiple context windows to address continuous task challenges.