docs: .mcpb format enables one-click MCP for non-devs; multi-agent 90% perf gains documented

[claude-yolo]

Summary

• Desktop Extensions go .mcpb: Anthropic published the engineering post on one-click MCP installation for Claude Desktop. The format is renamed .dxt → .mcpb (MCP Bundle). Bundled runtimes, automatic updates, OS keychain secrets. Non-technical users can now install MCP servers. Big deal for MCP adoption.
• Multi-agent Research: production war stories: Detailed post on how Anthropic built Claude Research. Key numbers: multi-agent outperformed single-agent by 90.2%, parallel tool calling cut research time 90%, agents use ~15× more tokens than chat. Practical lessons on orchestrator prompting, rainbow deployments, and LLM-as-judge evals.
• Agentic Misalignment: New research on LLMs as potential insider threats in agentic settings. Relevant for anyone building autonomous agents.

Key Changes

File	What
content/blog/engineering/desktop-extensions.md	NEW — .mcpb format spec, manifest schema, runtime bundling
content/blog/engineering/multi-agent-research-system.md	NEW — 8 prompting principles, parallel subagent architecture, eval strategies
content/blog/research/agentic-misalignment.md	NEW — insider threat research
content/blog/research/2028-ai-leadership.md	NEW — AI geopolitics scenarios
content/blog/research/81k-economics.md	NEW — 81k-person AI economics survey
content/blog/research/AI-assistance-coding-skills.md	NEW — AI impact on coding skill formation
content/github/.../marketplace.json	NVIDIA cuOpt skills plugin added
content/en/api/terraform/beta/**	HTML rebuild churn (CSS/JS hash rotation)

Noise filtered out

• 4 existing blog posts: redirect URL hash changes only
• claude-code-manifest.json: removed readmeFilename field, no version bump (still 2.1.159)

🤖 Generated by nightshift docs fetcher (claude-sonnet-4-6)