How Microsoft's AI VP Automates Everything with 5 Micro-Agent Workflows

In this episode of How I AI, Microsoft VP of Core AI Products and AI Futurist Marco Casalaina uses five small agent workflows for Azure permissions, document scanning, video compression, meeting scheduling, and content alerts.

Marco calls these ad hoc agents or micro-agents. They are narrow workflows created for one immediate problem rather than permanent systems with their own product surface.

The appeal is cumulative. Removing a few minutes of terminal lookup, file manipulation, or inbox work can also remove the context switch around the task.

Three workflows run through Warp on Marco’s computer. Two others use Microsoft 365 Copilot and ChatGPT to respond to an event or a schedule.

Workflow 1: Assign Azure roles with Warp

Marco uses Warp as a conversational interface to the Azure command-line tools. In the demonstration, he needed to assign several roles to a colleague on a subscription.

The problem it solves

Marco estimated that assigning each role through the Azure portal would have taken close to an hour.

The command workflow

He described the desired roles in plain language and let Warp construct the Azure CLI commands.

1. Step 1: Confirm the identity, subscription, role, and intended scope.
2. Step 2: Ask Warp to generate the commands. Role assignment changes cloud access, so inspect the target and scope before execution and record who approved it.

give him Azure AI user and Azure AI project Manager on this subscription that I'm looking at. ... actually I need to give him contributor role on the whole subscription

1. In the demonstration, Warp recovered from a failed command and completed the assignments. Marco described the run as taking seconds rather than the hour he estimated for the portal workflow.

Use rules and authoritative documentation

Marco gives Warp a rule that reminds him to activate owner access when required, and he connects it to Microsoft documentation for unfamiliar role names. Rules reduce repeated setup errors, but they do not decide whether a broad role such as subscription Contributor is appropriate. Least privilege remains the operator’s responsibility.

Workflow 2: Scan and merge a two-sided document

Marco needed to scan both sides of his daughter’s math practice test, then interleave the pages into one PDF.

The command workflow

He controlled the scanner from Warp rather than opening the manufacturer’s software.

1. Step 1: Install NAPS2. Marco configured a Warp rule with the local path and the feeder option for this open-source scanner tool.
2. Step 2: Put the first side in the feeder and ask Warp to scan it to a named file.

scan the documents from the feeder and save it to this directory as this file name

Warp invoked the scanner and saved the first PDF.

1. Step 3: Flip the pages, recall the previous command, and change the output from odd.pdf to even.pdf. This second pass did not require another model call.
2. Step 4: Ask Warp to merge the two files in page order.

Now put together the odd pages and the even pages and just make the math practice test out of it.

Warp wrote a temporary Python script using PyPDF2, interleaved the pages, created the combined document, and removed the script. Keep the source scans until the merged page order and legibility are verified, especially for records that cannot be recreated.

While the scanner ran, Marco returned to helping his daughter with the math problems. That small transfer of attention was the point of the workflow.

Workflow 3: Diagnose and compress a video

A ten-minute screen recording on Marco’s computer was 1.7 GB. He asked Warp to explain the size and re-encode the file at 1080p.

The prompt

The request combined diagnosis with a target outcome rather than prescribing every FFmpeg setting.

Why is this file so big use FM Peg to re-encode it? Still keeping it at 10 80 p... and make it more normal size.

The demonstrated result

Warp inspected the metadata, attributed the size to a very high bitrate and resolution, and ran FFmpeg. Marco reported that the output fell from 1.7 GB to 13 MB. That result is specific to this screen recording and encoding choice, so verify visual quality and audio before deleting the original.

Workflow 4: Trigger a meeting scheduler

In Microsoft 365 Copilot, Marco used the Workflows agent to build an Outlook automation for meeting requests from Claire.

The trigger and action

The instruction identifies a sender, extracts a proposed time, checks the calendar, and creates an invitation when the slot is free.

When I get an email from CLA VO requesting a meeting at a certain time, check my calendar. If that time is free, send her a 30 minute meeting invite for that time.

The saved workflow turns an incoming email into a calendar action without waiting for inbox triage. A durable version should allowlist the sender, require an explicit date and time zone, check conflicts and working hours, and avoid acting on forwarded or ambiguous requests.

Workflow 5: Schedule a content alert

Marco used a recurring ChatGPT task as a lightweight monitor for new How I AI episodes.

The prompt

He asked ChatGPT to check once a day and notify him when a new episode appeared.

every day look to see if there's a new podcast by CLA VO and notify me if there's a new one.

The demonstrated result

ChatGPT selected 9:00 AM for the daily check and offered desktop notifications. A monitor like this should link back to the source and distinguish a failed check from a confirmed no-update result.

Where micro-agents fit

The five examples share a narrow trigger and a visible outcome. Some are disposable terminal sessions. Others persist because a schedule or event makes the task recur.

Marco also uses AutoHotkey abbreviations for repeatable prompts. Typing MBF5 expands into an instruction to answer from Microsoft’s perspective in 500 characters or fewer with no bullets or formatting. He keeps a separate Warp rule against committing environment files that may contain secrets.

The decision is whether the task benefits from a saved trigger. Keep a rule for stable setup, save a recurring workflow when the event is reliable, and let one-off file operations remain disposable. For commands that change access, delete files, or contact people, keep the authorization boundary explicit.