AI News Summary - September 28, 2026

Claude Code Introduces Multi-Repository Context Window

Anthropic announced that Claude Code now supports simultaneous context from up to 10 linked repositories, with a combined context window of 500K tokens. The feature uses intelligent caching to maintain conversation continuity across codebases while reducing API costs by 60%. Early adopters report significant improvements in refactoring tasks that span multiple microservices.

MCP 3.0 Specification Released with Bidirectional Tool Calling

The Model Context Protocol consortium unveiled MCP 3.0, introducing bidirectional tool calling that allows servers to request actions from clients during execution. This enables more sophisticated agent workflows where AI assistants can negotiate permissions and request human input mid-task. Over 40 companies including Google, Microsoft, and Anthropic have committed to implementing the new spec within Q4 2026.

Open-Source Agent Framework "Orchestrator" Reaches 50K GitHub Stars

The Orchestrator project, which simplifies building multi-agent systems with built-in fault tolerance and state management, has become one of the fastest-growing AI frameworks. Version 2.0 adds native support for MCP, automatic prompt optimization, and a visual debugging interface. The framework now powers agent deployments at over 2,000 organizations.

Cursor's New "Team Memory" Feature Learns Organization Patterns

Cursor rolled out Team Memory, which automatically learns and enforces team-specific coding patterns, architectural decisions, and style preferences across all developers. The system builds a shared knowledge graph from merged PRs and code reviews, reducing onboarding time for new developers by an average of 40%. Enterprise pricing starts at $50/user/month.