How I AIHow Microsoft's AI VP automates everything with Warp | Marco Casalaina
CHAPTERS
- 0:00 – 2:30
Marco’s core idea: “micro-agents” that remove everyday friction
Marco frames Warp as a surprisingly powerful way to automate small, real-world tasks—often outside of coding. Claire sets the theme: lightweight agents can be one-off (ad hoc) or recurring/triggered, all aimed at reducing operational friction.
- •Warp as “pretty magical,” and even better with a bit of setup
- •Concept of micro-agents for small tasks (one-off vs recurring)
- •Why terminal/CLI + AI is a high-leverage combo
- •Episode focus: fast demos of practical automations
- 2:30 – 3:58
Why Marco chose Warp for administrative work (not just coding)
Marco explains he adopted Warp after Microsoft’s PowerShell team recommended it. He quickly realized Warp’s real value for him was automating admin tasks via command-line tools, especially cloud management.
- •Introduced to Warp by Microsoft’s PowerShell team
- •Hooked after using it for Azure subscription and access tasks
- •Uses Warp less as a coding agent and more as a CLI automation layer
- •CLI availability is the key criterion for what Warp can automate
- 3:58 – 6:01
Demo: Assigning Azure roles and permissions via Warp + Azure CLI
Marco shows how Warp can grant a colleague multiple Azure roles quickly using the Azure CLI (az). What would take a long time through the Azure portal becomes a short conversational request with iterative command execution and error recovery.
- •Granular role assignment (Azure AI User, Project Manager, Contributor)
- •Warp repeatedly calls az commands to complete the workflow
- •Agent recovers from command errors and continues to completion
- •Massive time savings vs clicking through the web portal UI
- 6:01 – 7:54
Why CLI-first workflows beat complex GUIs for cloud/IAM tasks
Claire and Marco discuss why cloud permission UIs are inherently difficult to design and use. They argue that AI paired with CLIs/APIs can abstract away the front-end complexity and keep users in a single, efficient loop.
- •GUI design for permissions/configuration is a “hard design problem”
- •AI access to CLIs/APIs/MCPs provides a simpler interface layer
- •Replaces the old search-copy-paste-debug loop (e.g., Stack Overflow flow)
- •Terminal becomes the unified place to ask, run, fix, and rerun
- 7:54 – 10:57
Improving reliability: MCP servers + simple “rules” in Warp
Marco explains the two main levers that make Warp perform better: connecting to an MCP documentation server and setting lightweight rules. These guardrails help Warp choose correct roles and avoid common failure modes (like missing owner activation).
- •Connect Warp to Microsoft Docs MCP server for Azure role discovery
- •Use rules to prevent repeated mistakes and missing prerequisites
- •Example rule: remind Marco to activate owner access before role changes
- •Rules don’t need to be fancy—just pragmatic, conversational constraints
- 10:57 – 13:02
Demo: One-command document scanning automation (homework, keepsakes)
Marco demonstrates using Warp to trigger a physical scanner via command line, saving directly to a chosen directory and filename. Claire highlights how this removes painful native scanning software steps and turns scanning into a quick, repeatable workflow.
- •Warp triggers the scanner and scans to a specified path/filename
- •Minimal manual work: load feeder, run command, optionally flip pages
- •Use case: scanning practice tests to generate problem variants in ChatGPT
- •Personal archiving: scanning cards and kids’ keepsakes for safety
- 13:02 – 15:35
Combining odd/even scans: Warp’s “bimodal” flow + quick Python glue
After scanning one side, Marco uses Warp’s command-history behavior to rerun and tweak the command for even pages. He then asks Warp to merge PDFs, and Warp generates and runs a short Python script (installing PyPDF2 if needed) and cleans up afterward.
- •Warp can operate as agent and as direct command execution (bimodal)
- •Up-arrow recalls generated CLI command for fast reruns with edits
- •Warp writes/runs a temporary Python script to merge PDFs correctly
- •Automates tedious PDF manipulation that would be manual in GUIs
- 15:35 – 17:40
It’s not pure magic: installing the right CLI tools (NAPS2)
Marco notes the scanning automation depends on having a CLI-capable scanning tool installed. He uses NAPS2 on Windows and encodes the tool path and scanning switches into Warp rules so it works consistently every time.
- •Windows scanning via CLI typically requires extra tooling
- •NAPS2 provides an open-source scanner CLI for Windows
- •Rule includes executable path and correct switches (feeder vs flatbed)
- •Preparation enables repeatable automation and fewer agent failures
- 17:40 – 20:23
Demo: Video compression and file diagnostics with FFmpeg
Marco shows Warp diagnosing why a screen recording is enormous and then re-encoding it using FFmpeg while preserving 1080p. The example demonstrates file introspection plus correct CLI invocation to shrink a 1.7GB video to a manageable size.
- •Warp analyzes file properties (bitrate/resolution) to explain size
- •Uses FFmpeg CLI to re-encode while keeping target resolution
- •Example outcome: 1.7GB → ~13MB for a 10-minute screen recording
- •FFmpeg also used for audio fixes (e.g., boosting volume in a time range)
- 20:23 – 22:51
The mindset shift: “ad hoc agents” and embracing ephemeral solutions
Marco calls these one-off, on-the-fly automations “ad hoc agents,” a pattern emerging across tools. Claire argues many of these solutions should remain disposable—redo them later with better models—while saving only lightweight rules when needed.
- •Ad hoc agents = unnamed mini-agents created on demand for a task
- •Trend: general-purpose agents increasingly support this pattern
- •Prefer ephemeral solutions over over-engineering production tools
- •Save rules/guardrails only when repetition or consistency is needed
- 22:51 – 26:06
Triggered workflows in Microsoft 365 Copilot (agents that build agents)
Marco demonstrates M365 Copilot’s workflow builder creating an email-triggered automation. The workflow extracts meeting times from Claire’s email, checks Marco’s calendar, and sends a 30-minute invite automatically if he’s free.
- •Copilot evolves from Q&A to agent building (builder builds an agent)
- •Trigger: email from a specific sender requesting a meeting time
- •Extracts time into ISO 8601 for Outlook API compatibility
- •Checks calendar availability, then auto-creates/sends meeting invite
- 26:06 – 27:48
Scheduled tasks in ChatGPT: cron-like recurring personal automations
Marco shows similar automation in a consumer tool: ChatGPT can run a recurring check for new podcast episodes and notify him. Claire frames this as shifting tasks off the human critical path to stay responsive without constant monitoring.
- •Recurring trigger: daily check for new Claire Vo podcast episodes
- •ChatGPT chooses a default run time (e.g., 9 AM) and can notify desktop
- •Demonstrates parity between enterprise and consumer agent features
- •“Anti-to-do list”: remove yourself from the loop for routine vigilance
- 27:48 – 34:09
Time management, teaching kids, and debugging AI with prompt shortcuts
In a lightning round, Marco explains these automations save minutes daily and enable parallel work (e.g., scanning while tutoring). He notes his daughter is a capable but mainstream AI user, and he shares how he boosts reliability with rules and AutoHotkey prompt snippets.
- •Practical impact: multitask while agents handle admin/scanning chores
- •Kids adoption: daughter prefers mainstream tools (e.g., Canva) over tinkering
- •When AI fails: add rules (e.g., never commit .env) and constrain outputs
- •AutoHotkey library of reusable prompt snippets (e.g., 500-char response)
