Anthropic launches Managed Agents

Artificial Analysis

GPT-5.4 tops the APEX-Agents-AA benchmark for AI agents at 33.3%, slightly ahead of Claude Opus 4.6 and Gemini 3.1 Pro, though all leading models score below 35% on these long-horizon professional tasks. The results highlight that agentic AI capabilities remain challenging even for frontier models.

Anthropic

Anthropic's Managed Agents is a hosted service that decouples AI agent components (brain, hands, session) into stable interfaces, enabling reliable long-horizon task execution while dramatically improving performance and allowing flexible deployment to customer infrastructure.

arXiv

Meta AI's SandMLE framework enables scalable reinforcement learning for ML engineering agents by using synthetic micro-scale datasets, reducing training time by 13x while achieving 20-67% performance improvements over supervised fine-tuning approaches. This breakthrough makes on-policy RL practical for training agents that can handle complex machine learning workflows beyond basic software engineering tasks.

LangChain

LangChain presents a practical framework for self-improving AI agents that uses evals as training data to systematically hill-climb agent performance through iterative harness improvements. The company is releasing tooling to enable teams to build autonomous agent improvement systems with proper evaluation guardrails.

Hugging Face

Z.ai's GLM-5.1 matches Claude Opus coding performance at lower cost, excelling at long-horizon agentic tasks by sustaining optimization through hundreds of iterations rather than plateauing early like previous models.