OpenHands 1.32 lets the MCP install dialog choose between several installable library options and lists repositories from every cloneable provider in the automation picker. Pinned automations are now stored in agent-server settings, and canvas backend logs are quieter by default.
Codex 0.164 lets users save prompts and slash commands in ten numbered /presets slots and set model, reasoning effort and permissions directly from /model and /permissions. It also adds enterprise MCP authentication, credential masking in remote shell sessions and an Amazon Bedrock setup wizard turned on by default.
The Agentic AI Foundation, which maintains open-source AI tools, has taken on four new projects donated by OpenAI, Anthropic, Hugging Face and Alibaba. That brings the number of projects it maintains to 10.
A Microsoft post on the Hugging Face blog presents ThinkingBox, an approach that turns database grading into reinforcement learning rewards for post-training agents.
An MIT Technology Review piece argues that the next edge in enterprise AI lies less in picking a model than in what gets built around it. After years focused on bigger models and benchmark scores, it says the surrounding system now matters more.
Replit has launched Monitoring Agent, an autonomous on-call engineer for apps built on its platform. When a production issue hits, it investigates the alert, finds the cause, makes and validates a fix, deploys it and reports back on what changed.