A curated roundup of AI/ML news from July 3–18, 2026 covering five major developments: OpenAI's GPT-5.6 GA with programmatic tool calling and multi-agent orchestration; PyTorch 2.13's training and inference optimizations including fused loss computation and FlexAttention on Apple Silicon; Meta's Muse Spark 1.1 and SpaceXAI's Grok 4.5 competing in the agentic API market; Google's Gemini Managed Agents gaining background/async execution with persistent state; and the GhostApproval vulnerability affecting Amazon Q Developer, Claude Code, Cursor, and others via symlink attacks that bypassed human approval. The overarching theme: orchestration is becoming a platform primitive, agent harnesses matter as much as models, and human-in-the-loop approval is only meaningful when the system shows accurate provenance.
