How Microsoft's AI VP automates everything with Warp | Marco Casalaina
TL;DR
Marco Casalaina uses Warp less like a coding IDE and more like an ad hoc agent layer over CLIs — the Microsoft VP says he got hooked when Warp started reliably handling Azure admin tasks like assigning Azure AI User, Azure AI Project Manager, and Contributor roles through the Azure CLI instead of an hour of painful portal clicking.
The real unlock is pairing AI with rules and MCP servers, not perfect prompting — Casalaina improves Warp by connecting Microsoft Docs via MCP for role lookup and adding simple conversational rules like “remind me to activate owner access first,” which makes Azure workflows much more reliable.
He applies the same pattern to weirdly practical home tasks, including scanning his daughter’s two-sided math practice test — with an open-source Windows scanner CLI called NAPS2 and a Warp rule pointing to the install path, he can start scans from the feeder, merge odd/even pages, and generate a finished PDF without touching the scanner software.
Warp also acts as a file-manipulation agent for media work — after Xbox Game Bar produced a 1.7 GB file for a 10-minute screen recording, he asked Warp to inspect it and re-encode with FFmpeg at 1080p, cutting it down to 13 MB while explaining the oversized bitrate and resolution.
Casalaina’s bigger thesis is that ‘micro agents’ are everywhere now: one-off, triggered, and recurring — he demos Microsoft 365 Copilot building an email-triggered workflow that auto-schedules meetings from Clarvo’s email requests, then shows ChatGPT creating a daily 9:00 a.m. check for new podcast episodes like a cron job for consumer AI.
His operating style is lightweight and highly repeatable, not precious — when AI misbehaves, he adds small rules or reusable AutoHotkey prompt snippets like “answer from the perspective of Microsoft in 500 characters or less,” rather than building heavyweight systems or obsessing over formal prompt design.
The Breakdown
Why Warp clicked for a Microsoft AI VP
Marco Casalaina says he started using Warp because Microsoft’s own PowerShell team nudged him toward it, and he got hooked once it proved itself on Azure administration. The funny premise of the episode is that “the sexiest” AI workflow here is role assignment and infra admin — but that’s exactly the point: these annoying DevOps chores eat the time that should go to actual building.
Azure role assignment without the portal pain
His first demo is giving a colleague, Govind, access to Azure resources by assigning Azure AI User, Azure AI Project Manager, and later Contributor on a subscription. Instead of hunting through Azure’s web UI role by role, he prompts Warp, which keeps hammering the Azure CLI until it gets the job done — including recovering from a mistake mid-run. Casalaina’s blunt takeaway: whenever there’s a usable CLI, Warp is “freaking great at that.”
The secret sauce: docs MCPs and plain-English rules
He’s clear that the magic gets better with a little setup. For Azure work, he sometimes connects Warp to the Microsoft Docs MCP server so the agent can look up which role is actually needed for something like Azure Document Intelligence, instead of guessing. He also adds simple rules — like reminding him to activate owner access before assigning resource group roles — and Clarvo points out how refreshingly un-fancy the prompting is: just a few practical instructions in conversational English.
Scanning homework like a command-line gremlin
Then the episode gets wonderfully domestic. Marco scans his daughter’s two-sided math practice test by telling Warp to start the scanner, save odd pages, then scan even pages — and yes, Clarvo stops to confirm that this literally wakes the scanner up remotely. To stitch it together, Warp installs PyPDF2, writes a quick Python script, merges the files, and deletes the script afterward; Casalaina jokes that this is “using Warp as a coding agent sideways.”
It’s not magic-magic: NAPS2 made the trick possible
Casalaina uses the scanner example to make an important point: agentic workflows often depend on finding the right command-line surface first. On Windows, he installed the open-source NAPS2 scanner CLI and added a Warp rule telling it where NAPS2 lives and which switch uses the feeder instead of the flatbed. Once that rule existed, he says, Warp stopped messing it up and got it right every time.
FFmpeg, giant files, and underappreciated file manipulation
For video, he shows a 10-minute Xbox Game Bar screen capture that somehow ballooned to 1.7 GB. He asks Warp why it’s so big and tells it to re-encode with FFmpeg while preserving 1080p; Warp inspects the file, identifies an absurd bitrate/resolution combo, and spits out a 13 MB version. Clarvo uses this to argue that file manipulation — not just file generation — is one of AI’s most underrated powers.
From one-off terminal tasks to recurring micro agents
Marco frames all of this as creating “ad hoc agents” on the fly: unnamed little workers for specific jobs. Then he broadens the idea with Microsoft 365 Copilot, showing a workflow that turns an email from Clarvo asking for a meeting into a triggered agent that extracts the time in ISO 8601, checks Outlook, and sends a 30-minute invite if he’s free. After that, he shows the same pattern in ChatGPT: a daily 9:00 a.m. recurring task that checks for new Clarvo podcast episodes and can push desktop notifications.
What this changes in real life
In the lightning round, Casalaina says these tools save him “many minutes a day,” but the deeper win is parallelism — while Warp scans homework, he and his daughter can work the math problems instead of fighting terrible scanner software. His daughter isn’t the tinkerer he is, though he says she’s a “wizard at Canva,” and his own style is very pragmatic: if AI fails, he adds a rule, and for repeatable writing tasks he uses artisanal AutoHotkey shortcuts to expand prompts like “answer from the perspective of Microsoft in 500 characters or less.”