The skills vs MCP distinction is where I think most teams get stuck. Skills are lighter to iterate on (markdown, no runtime dependencies), but MCP gives you real-time state. Curious — when you're building a new agent workflow, do you default to skills first and only reach for MCP when you hit a wall, or do you pre-decide based on whether the tool needs live connectivity?
Also — have you tried any non-Google MCP servers that surprised you? Most of the ecosystem is still clustered around OpenAI/Anthropic tooling, but some of the community-built servers (filesystem, browser-use, postgres) are getting genuinely good.