copilot-studio-agent-mastery-projects

Microsoft Copilot Studio — Progressive Agent Mastery Portfolio (2026)

A component-by-component Copilot Studio project portfolio: nine hands-on projects spanning Beginner → Intermediate → Advanced → Expert, plus a capstone that unifies every capability into one governed enterprise mesh. Each project isolates one Copilot Studio building block — knowledge grounding, web/anonymous publishing, Power Automate actions, structured Dataverse/SharePoint data, custom APIs & MCP servers, generative orchestration, multi-agent systems, and governance — so you can learn (and demonstrate) each piece deliberately instead of copying one tutorial that vaguely touches everything.

Kept current as of July 4, 2026. Copilot Studio’s licensing model changed materially in the last year (messages → Copilot Credits, September 2025; Agent Pre-Purchase Plan added February 2026; multi-agent orchestration GA April 2026; New Orchestrator/Agentic Reasoning Loop and computer-using agents GA May 2026). Every project below reflects that current state — see the licensing primer below before you build anything.

📂 Projects by difficulty

Tier	#	Project	Core capability	Preview
🟢 Beginner	1	KB-Agent-101	Knowledge sources, generative answers	🔗 Preview
🟢 Beginner	2	WebGround-Agent	Public website grounding, anonymous channel	🔗 Preview
🟡 Intermediate	3	FlowConnect-Agent	Power Automate actions, agent flows, Adaptive Cards	🔗 Preview
🟡 Intermediate	4	DataLink-Agent	Dataverse MCP tool, SharePoint lists, code interpreter	🔗 Preview
🟠 Advanced	5	APIBridge-Agent	Custom connectors, self-hosted custom MCP server	🔗 Preview
🟠 Advanced	6	OrchestrAI-Agent	New Orchestrator, Skills, Bring Your Own Model	🔗 Preview
🔴 Expert	7	MultiAgentMesh	Embedded/connected agents, A2A interoperability, Fabric	🔗 Preview
🔴 Expert	8	GovernOps-Agent	Admin governance, credit monitoring, an agent that governs agents	🔗 Preview
⭐ Capstone	9	Unified Copilot Mesh	All eight capabilities, one governed system	🔗 Preview

Live previews are static HTML/CSS mockups approximating each project’s real Copilot Studio look & feel (chat UI, activity map, orchestration canvas, admin dashboards). They are visual references only, not functional agents.

🧭 How to use this repo

New to Copilot Studio? Build Projects 1 → 8 in order. Each one deliberately depends on lessons from the previous tier — Project 3’s action-safety pattern assumes you understand Project 1’s grounding; Project 7’s multi-agent security assumes Project 4’s data-security lessons.
Want one project to demo to leadership? Go straight to Project 9 (Capstone) — it references every pattern below it and closes on the cost/governance story that actually gets budget approved.
Only care about one capability? Every project is self-contained enough to build independently if you already understand the prerequisites called out in its README.

💳 Copilot Credits & Licensing Primer (current as of July 2026)

Read this before building anything — it’s the single biggest source of surprise bills and the thing every project in this repo explicitly reasons about.

The core model

Since September 1, 2025, Copilot Studio’s metered unit is the Copilot Credit (it replaced the older “messages” unit — same underlying meter, new name and framing).
A Copilot Credit measures the work an agent does per turn — retrieving information, grounding on knowledge, calling a tool/action, running a flow, or invoking generative/reasoning AI. Different activities cost different numbers of credits, and a single turn can stack multiple charges (e.g., a grounded answer that also triggers an action might cost 2 + 5 = 7 credits).
Rough reference rates used throughout this repo’s projects: generative answer ≈ 2 credits; tenant-graph grounding ≈ 10 credits; a standard action/tool call ≈ 5 credits; premium/advanced-reasoning tools cost meaningfully more; autonomous/proactive triggers cost a flat 25 credits — even for Microsoft 365 Copilot licensed users, since autonomous triggers are never zero-rated.

What’s free vs. what’s metered

If your users hold Microsoft 365 Copilot licenses ($21/user/month standard, promotional $18/user/month for ≤300 users through June 30, 2026 for Business; $30/user/month for Enterprise) and the agent runs inside Microsoft 365 (Copilot Chat, Teams, SharePoint) for authenticated, licensed users only — usage is largely zero-rated, subject to fair-use limits. This covers classic answers, generative answers, tenant-graph grounding, agent actions, and agent flows.
The moment an agent is published to an external/anonymous channel (public website, WhatsApp, a custom app) or used by unlicensed users, it requires a standalone Copilot Studio license and consumes metered Copilot Credits regardless of your M365 Copilot entitlements. This exact boundary is the focus of Project 2.

The four ways to pay for Copilot Credits

Pay-as-you-go — $0.01/credit, billed monthly through a linked Azure subscription. No upfront commitment; best for unpredictable or pilot-stage usage.
Prepaid Copilot Credit capacity packs — $200/month for 25,000 credits. Unused credits do not roll over. Best for predictable, steady volume.
Copilot Credit Commit Units (CCCU) pre-purchase plan — a one-year upfront pool; each CCCU = $1 of consumption (100 credits), with volume discounts from 5% at 3,000 CCCUs up to 20% at 3,000,000 CCCUs. Falls back to PAYG automatically once exhausted.
Microsoft Agent Pre-Purchase Plan (Agent Commit Units / ACUs) — added February 2026; works like CCCUs but draws down across Copilot Studio, Microsoft Foundry, and increasingly Fabric and GitHub agentic services, for organizations standardizing spend across Microsoft’s whole agentic stack.

Things every project in this repo explicitly accounts for

Bring Your Own Model (Foundry-hosted) usage incurs Copilot Credits for the Copilot Studio orchestration and separate Azure OpenAI/Foundry per-token charges for the model itself — two invoices, not one (see Project 6).
Managed Environments, if enabled, require every active user to be covered by a proper license or PAYG meter — limited Power Platform rights bundled in Dynamics 365/M365 licenses do not satisfy this (see Project 8).
AI Builder credits are being folded into the Copilot Credit model — seeded AI Builder credits end and Power Apps/Power Automate AI Builder usage shifts to Copilot Credit billing rates. Budget for this transition if your agents touch AI Builder-adjacent capabilities (see Projects 3, 4, and the Capstone).
Credits don’t roll over on prepaid capacity packs — don’t over-buy against optimistic volume forecasts.
Use the official Copilot Studio agent usage estimator (agent type, traffic, orchestration mode, knowledge, tools) to forecast cost before you scale traffic — every project’s “Token / Copilot Credit utilization” section in this repo is designed to feed directly into that estimator.

(Prices and mechanisms above are US list rates as published mid-2026; Microsoft updates these regularly — always confirm current figures on Microsoft’s live Copilot Studio pricing and licensing pages before committing budget.)

🏗️ Common practices applied across every project

Confirm-before-act on any state-changing action (Projects 3, 4, 6, 7) — no agent in this repo silently executes a write, a cancellation, or an approval without checking with a human first.
Security inherits from the data layer, not the LLM — Dataverse security roles and SharePoint permissions are the enforcement mechanism (Project 4); the agent never becomes the thing deciding who can see what.
External/non-Microsoft integrations carry explicit accountability documentation — Project 5 and 7 both call out, in writing, who owns vetting an external MCP server or A2A partner agent.
Cost is a design input, not an afterthought — every project’s architecture includes at least one decision (agent flow vs. cloud flow action, embedded vs. connected agent, deep reasoning on/off) made partly because of its Copilot Credit implications.
Governance scales with capability — the more autonomous and powerful an agent becomes (Project 6’s Agentic Reasoning Loop, Project 7’s multi-agent mesh), the more explicit its guardrails and consumption caps must be.

🔧 Platform & tooling baseline (as of July 2026)

Copilot Studio on the Copilot Credits consumption model (replacing “messages” since September 2025)
New Orchestrator / Agentic Reasoning Loop generally available (multi-step autonomous planning within a single agent)
Multi-agent orchestration generally available (April 2026) — embedded/child agents, connected agents, and Agent2Agent (A2A) interoperability with non-Copilot-Studio agents
Model Context Protocol (MCP) generally available for both consuming external MCP servers and publishing your own
Immersive Prompt Builder (GA) embedded directly in the agent’s Tools tab
Bring Your Own Model via deep Microsoft Foundry integration (11,000+ models, including GPT, Llama, DeepSeek, and fine-tuned custom models)
Code interpreter (GA) for Python-powered data analysis directly inside agent conversations
Fabric Data Agents for governed, audited grounding on enterprise analytics data
Microsoft Agent Pre-Purchase Plan (ACUs), added February 2026, spanning Copilot Studio, Foundry, and (increasingly) Fabric/GitHub agentic services
AI-powered governance agents in the Power Platform admin center for tenant-wide real-time risk assessment

📌 Suggested presentation order

Start with the Capstone (Project 9) for a manager/leadership audience — one coherent story across every difficulty tier, closing on cost governance.
Use Projects 1-8 as deep-dive appendices when a specific stakeholder (security, finance, a particular business team) wants to go deeper on exactly one capability.

📄 License

This repository is provided as a personal learning/portfolio and internal proof-of-concept reference. Adapt freely for your own organization’s Copilot Studio POC and training needs. Licensing and pricing figures cited throughout are informational summaries of Microsoft’s published rates as of mid-2026 — always verify current terms on Microsoft’s official Copilot Studio licensing and pricing pages before making purchasing decisions.