AI News - September 7, 2026

GitHub Copilot Workspace Launches Multi-Agent Collaboration Mode

GitHub announced Copilot Workspace now supports multiple specialized AI agents working simultaneously on different parts of a codebase. The feature allows developers to assign separate agents for testing, documentation, and implementation tasks that coordinate through a shared context window. Early beta testers report 40% faster completion times for complex refactoring projects.

Anthropic Releases MCP 2.1 with Bidirectional Tool Calling

The Model Context Protocol specification has been updated to version 2.1, introducing bidirectional tool invocation that allows servers to request capabilities from connected clients. This enables more sophisticated workflows where AI assistants can negotiate resource access and coordinate across multiple MCP servers. Over 200 MCP server implementations have been published to the community registry this month.

Cursor Introduces "Code Memory" for Project-Wide Learning

Cursor's latest update adds persistent memory that learns from your codebase patterns, team conventions, and past corrections over time. The system builds a private knowledge graph specific to each project, reducing repetitive explanations and improving suggestion accuracy. The feature runs entirely locally and encrypts all learned patterns.

OpenHands Agent Framework Reaches 1.0 with Docker Isolation

The open-source OpenHands autonomous coding agent has hit its 1.0 milestone with production-ready Docker sandboxing for safe code execution. The framework now includes built-in security policies, resource limits, and a plugin system for extending agent capabilities. Initial benchmarks show it can complete 68% of SWE-bench Verified tasks autonomously.