Скачать или смотреть Building AI Agents in PowerShell: GitHub CLI as an Autonomous Workflow

In this livestream, I show how to build AI agents in PowerShell by turning the GitHub CLI into an autonomous workflow engine.

No frameworks. No mythology.

We wrap GitHub CLI commands in PowerShell functions, expose them as tools to an LLM, and let the model infer which actions to take from plain English. Creating issues, closing issues, assigning Copilot, switching models, and running everything headless from the CLI.

Topics covered:

Building AI agents from scratch in PowerShell
Function calling and tool schemas
Using GitHub CLI as an automation surface
Model choice, failures, and why evals matter
Running agents locally, in GitHub Actions, or in the cloud
Why agents are just tools + prompts + a loop

This isn’t about GitHub specifically. GitHub is just a good demo. The same pattern works for any CLI: Azure, AWS, Jira, internal tooling, whatever.

If you want to understand how agent frameworks actually work under the hood, this is the machinery.

Here’s a *clean, solid YouTube ToC* that matches how the talk actually flows, not how marketing decks pretend it flowed.

Table of Contents (YouTube Chapters)

00:00 – Intro: What We're Building Today
02:10 – Why Agents Aren't Magic
05:05 – Setup: PowerShell, PSAI, API Keys
08:40 – The Agent Mental Model (Tools, Intent, Loop)

12:15 – Wrapping GitHub CLI as PowerShell Functions
16:30 – Exposing Functions as Agent Tools

19:24 – Agents Are Just Functions + Intent
21:09 – Why This Pattern Scales Beyond GitHub

24:10 – Running the First End-to-End Agent Workflow
28:45 – Tool Selection and Parameter Inference

33:39 – The Model Messed Up… So I Swapped the Brain
38:24 – Why Model Choice and Evals Matter

41:10 – Guardrails, Safety, and "Don't Nuke Your Repo"
49:09 – "Close All Issues"
49:49 – Why Agents Need Constraints

50:24 – Copilot Is Just Another Agent
52:24 – Thinking in Primitives, Not Products

55:10 – Where This Runs: Local, CI, Cloud
58:40 – Generalizing the Pattern to Any CLI
01:02:30 – Final Takeaways: How to Build Agents for Real
01:05:00 – Wrap-Up and Q&A
Table of Contents (YouTube Chapters)

00:00 – Intro: What We're Building Today
02:10 – Why Agents Aren't Magic
05:05 – Setup: PowerShell, PSAI, API Keys
08:40 – The Agent Mental Model (Tools, Intent, Loop)

12:15 – Wrapping GitHub CLI as PowerShell Functions
16:30 – Exposing Functions as Agent Tools

19:24 – Agents Are Just Functions + Intent
21:09 – Why This Pattern Scales Beyond GitHub

24:10 – Running the First End-to-End Agent Workflow
28:45 – Tool Selection and Parameter Inference

33:39 – The Model Messed Up… So I Swapped the Brain
38:24 – Why Model Choice and Evals Matter

41:10 – Guardrails, Safety, and “Don't Nuke Your Repo”
49:09 – “Close All Issues” (The Oh-Shit Moment)
49:49 – Why Agents Need Constraints

50:24 – Copilot Is Just Another Agent
52:24 – Thinking in Primitives, Not Products

55:10 – Where This Runs: Local, CI, Cloud
58:40 – Generalizing the Pattern to Any CLI
01:02:30 – Final Takeaways: How to Build Agents for Real

#PowerShell, #AIAgents, #GitHubCLI, #Automation, #LLMs, #DevTools, #AgenticWorkflows