AI News Brief - September 25, 2026

Cursor IDE Introduces Multi-Agent Code Review System

Cursor has launched "Review Swarm," a feature that deploys multiple specialized AI agents to analyze pull requests from different perspectives (security, performance, accessibility). Early adopters report 40% fewer bugs reaching production compared to single-agent reviews. The system integrates with existing CI/CD pipelines and supports custom agent configurations.

MCP 2.3 Specification Adds Streaming Context Support

The Model Context Protocol consortium released version 2.3, introducing real-time streaming context updates that allow AI assistants to maintain live connections to data sources. Major implementations from Anthropic, Microsoft, and Google are expected within 30 days. The update reduces context staleness issues that plagued long-running agent sessions.

Open-Source AgentKit Reaches 1.0 Release

The collaborative agent framework AgentKit hit its first stable release after 18 months of development, offering production-ready tools for building multi-agent systems with built-in observability. The framework features native MCP support, automatic tool discovery, and agent-to-agent negotiation protocols. Over 200 organizations contributed to the project, which is now deployed in several Fortune 500 companies.

GitHub Copilot Workspace Adds Visual Code Generation

GitHub unveiled Copilot Workspace's new "Sketch-to-Code" capability that generates functional components from hand-drawn diagrams or wireframes. The feature uses multimodal models to interpret visual intent and produces production-ready code in React, Vue, or Svelte. Beta testers report 3x faster prototyping for UI-heavy features.