Diagnosing instability in production-scale agent reinforcement learning

Here's a 3-sentence summary of the blog post: To improve post-training outcomes for large-scale multimodal agents at Microsoft, the authors developed five engineering and algorithmic interventions that address silent failures, policy gradient estimator degradation, and aggregate reward climbing while gradient signal concentrates. These interventions include a staged objective curriculum, adaptive curriculum from estimator health, variance-corrected normalization from estimator structure, and two methods to counteract degenerate behaviors such as a KL penalty for entropy regularization and a reservoir buffer for restoring outcome contrast. By implementing these interventions, the authors successfully improved the performance and robustness of their multimodal agents, reducing silent failures and improving overall results at scale.

Here's a concise 3-sentence summary of the Microsoft Learn MCP Server engineering blog post: Microsoft developed the Learn MCP Server to enable AI agents like GitHub Copilot to access trusted, up-to-date Microsoft Learn documentation through a standardized Model Context Protocol (MCP). The remote server exposes agent-friendly tools over Streamable HTTP Transport, backed by the Learn knowledge service, allowing agents to ground responses just like humans with browsers do. The Learn team learned valuable lessons during the design, launch, and operation of the MCP Server, including the importance of designing tools for the agent workflow, hosting a public server like a distributed system, and using tool descriptions to improve agent experience.