Ready-to-run samples for Agents and Workflows in Copilot Studio. The retro-future game store, run by agents.
Copilot Studio has been rebuilt for complex, multi-step work — coordinated, multi-agent systems on the GitHub Copilot harness, where the agent reasons, acts, and adapts in a loop. The fastest way to understand it is to run it. This guide from the CAT team is built around complete, ready-to-run samples: deploy one into a Power Platform environment you already have and watch connected agents, MCP servers, runtime skills, and generated files work together end to end. Copilot Studio has been rebuilt for complex, multi-step work, moving from a single chatbot to coordinated, multi-agent systems on the GitHub Copilot harness — its agentic build, where the agent reasons, acts, and adapts in a continuous loop. BlastBox Omega is a complete, runnable reference: a parent agent that delegates to connected agents, calls multiple MCP servers, runs per-agent skills that generate and execute Python at runtime, and returns real generated files. End-to-end reference scenarios spanning agents and workflows in Copilot Studio.
Walkthroughs
End-to-end transcripts that put the building blocks to work, from a quick self-serve task to a full multi-agent flagship. More scenarios will land here over time.
A member lost their BlastPass card. The agent verifies their identity, reissues the card, and renders a fresh tier-colored membership card. Membership MCP + card skill.
A defective console trade-up: rule the warranty, check stock, prorate the refund, apply loyalty points, and settle the balance. Two connected agents, four MCPs, chained skills, settlement PDF.
A manager works out end-of-quarter markdowns by drilling from a store-wide view down to one SKU week-by-week, then to policy-grounded discounts. Three connected agents, live charts, recommendations, manager PDF.
Get started
One command downloads the sample and deploys all four agents and their MCP connectors into a Power Platform environment you already have — no clone, no servers, nothing to host. A few minutes later you have a running store to explore.
Needs the pac CLI (signed in) and az CLI. Prefer to clone? Grab the sample and run node deploy/deploy.mjs.
# macOS / Linux
curl -fsSL https://raw.githubusercontent.com/microsoft/new-copilot-studio-tech-guide/main/deploy/install.sh | bash# Windows · PowerShell
powershell -c "irm https://raw.githubusercontent.com/microsoft/new-copilot-studio-tech-guide/main/deploy/install.ps1 | iex"Pick your pac profile and target environment when it asks.
Imports both solutions, deploys the MCP connectors, creates connections, publishes the agents.
One ~2-minute manual UI step it prints at the end.
Open each agent's Preview and run the walkthroughs above.
Go deeper
Reference material to go further: a technical deep-dive deck, the open-source plugin that upgrades and builds agents, and straight answers to the common questions.
A slide-by-slide walkthrough of the architecture behind Copilot Studio agents, connected agents, MCP, skills, orchestration, and file generation.
An open-source plugin for AI coding agents that upgrades existing Copilot Studio agents to the GitHub Copilot harness, or builds new agents from scratch.
View on GitHubStraight answers on upgrading agents, building on the GHCP harness, and what happens to topics, prompts, adaptive cards, and child agents.
Read the FAQ