How Microsoft's AI VP automates everything with Warp | Marco Casalaina

What It Covers

Microsoft VP Marco Casalaina demonstrates five micro-agent workflows using Warp terminal, showing how CLI-based AI agents handle Azure role assignment, document scanning, video compression, calendar scheduling, and podcast monitoring — reducing administrative friction without requiring formal prompting or complex automation setup.

Key Questions Answered

• •CLI-First Agent Strategy: Warp excels at any task with a command-line interface, not just coding. Casalaina assigns multiple Azure IAM roles to colleagues in minutes by prompting conversationally, replacing a process that previously required navigating Azure's web portal individually for each role and could consume an hour of work.
• •Rules as Persistent Memory: Add Warp rules to encode recurring context — such as the exact file path to a scanner CLI tool (NAPS2) or a reminder to activate owner access before Azure operations. Once a rule is set, the agent executes correctly every subsequent time without rediscovering steps or repeating failures.
• •MCP Servers Extend Agent Accuracy: Connecting Warp to the Microsoft Docs MCP server eliminates guesswork when assigning unfamiliar Azure roles. Instead of specifying exact role names, users can describe the desired capability and the agent looks up the correct role from official documentation before executing the CLI commands.
• •Triggered Agents Replace Manual Scheduling: Microsoft 365 Copilot's Workflows agent builds email-triggered automations through plain-language prompts. Describing a rule — check calendar when a specific contact requests a meeting, then send an invite if free — generates a functional Outlook-connected agent without writing code or configuring traditional workflow tools.
• •Ephemeral Agents Over Saved Scripts: Rather than productionizing one-off automations, re-run them fresh each time using an improved model. Casalaina compressed a 1.7GB screen recording to 13MB via FFmpeg through a single Warp prompt, then discarded the generated Python script — treating the agent as disposable infrastructure rather than permanent code.