I tracked my workflow for a week last year. Twelve hours of actual coding. The rest was spent bouncing between Cosmos DB Data Explorer, Azure Monitor, half-stale internal docs, terminal sessions I forgot about, and Slack threads where someone solved my exact problem in March.

We keep optimizing for writing code faster. But writing code was never the bottleneck.

This talk takes a different angle: what happens when you wire up AI agents with real capabilities—MCP servers that expose live data, packaged skills that encode team knowledge, and CLI tools that agents can invoke and chain together—so developers stay in one place instead of context-switching across ten.

I'll demo this inside VS Code with real scenarios: an agent querying Cosmos DB through MCP to investigate a performance regression, a reusable skill that encodes our team's troubleshooting runbook so any developer can run it in natural language, and CLI orchestration where the agent composes az commands, runs them, and adapts based on output—without me writing a shell script.

Under the hood I'll walk through how MCP handles tool discovery and execution, how skills package domain expertise into portable, shareable units that agents can pick up contextually, and how CLI integration turns existing command-line tools into agent capabilities without rewriting anything.

I'll also be upfront about where things broke. MCP servers that exposed too much and confused the agent. Skills that worked great for one team and hallucinated for another. CLI chains that ran confidently in the wrong direction. Trust boundaries and scoping turned out to be the real engineering work.

If you're building developer platforms or thinking about where AI fits in your team's daily work, this is the pragmatic, lessons-learned version of that story.