Multi-agent orchestration in 2026 — MCP for tools, A2A for agent-to-agent, LangGraph for stateful flows. Reference architecture, picking criteria, and production patterns.
Inside Anthropic Cowork mode — desktop agent architecture, MCP plugin model, sandboxed shell, computer use tier system, and how it differs from Claude Code in 2026.
How Anthropic Claude Skills inject narrow, on-demand capabilities into production LLM agents without bloating the system prompt. Architecture, patterns, trade-offs.