Building Goal-Driven Agents in Microsoft Copilot Studio (GitHub Copilot Harness)
kod szkolenia: MS Copilot Studio_level 1_GH.Harness / ENG DL 2d
This course teaches you how to build agents to which you describe a goal, while the agent itself plans how to achieve it. Instead of topic trees and rigid conversation paths, you define Instructions, Knowledge, Tools, Skills, and the Model, and generative orchestration selects the next steps, knowledge, and tools – even when one of the steps fails. The greatest value lies in autonomy, reasoning, and flexible goal completion within safe boundaries (guardrails, permissions, DLP).
Participants work in the new authoring experience (New experience) with the Build, Preview, Evaluate, and Monitor tabs. The GitHub Copilot harness is the most capable Copilot Studio runtime, designed for reasoning-heavy, multi-step work – the agent plans the steps, selects knowledge and tools, and, when needed, performs specific actions in enterprise systems.
This is a proprietary course. Participants build an AI agent, ground it in SharePoint knowledge, extend it with tools (connectors, MCP) and reusable skills in SKILL.md files, automate processes with Workflows following the “the agent decides – the workflow executes” model, and finally publish the agent to Microsoft Teams and Microsoft 365 Copilot. Each module ends with a step-by-step lab in which the participant extends the IT Concierge agent – from a baseline in the Preview tab to measurable response quality in the Evaluate tab.
Choose this course when the outcome is known but the path depends on data, tools, and intermediate results: the agent must complete an entire end-to-end process, work with Office and PDF documents, use several tools at once, or rely on Skills, MCP, memory, or connected agents. It also works well where use cases are variable and hard to describe with a decision tree – with no coding, but with an understanding of architecture, orchestration, security, and cost (Copilot Credits).
This course is intended for:
- Power Platform developers and low-code makers, who want to design reasoning agents that integrate enterprise data and systems using tools, skills, and workflows.
- IT professionals and Microsoft 365 administrators, responsible for tenant configuration, permissions, publishing, and secure sharing of agents across the organization.
- Business analysts and process owners, who want to automate complex, multi-step processes and improve the quality of service for users and customers.
- AI solution consultants and architects, who design Copilot Studio implementations aligned with governance best practices, ALM, and cost control.
- People getting started with AI agents, who want to learn the latest, generative approach to building agents (New experience) right from the start.
After completing this course, participants will gain the following benefits:
- A complete “from idea to deployment” workshop – you will independently build, extend, test, and publish a working AI agent.
- Building agents in natural language – you will learn to describe the agent’s goal instead of designing topic trees – generative orchestration plans the rest.
- An agent grounded in enterprise knowledge (RAG / grounding) – you will ground responses in data from SharePoint and files, respecting user permissions and citing sources.
- Tools and skills – you will connect the agent to systems via connectors and MCP, and package repeatable behaviors as reusable skills – available only in the GitHub Copilot harness.
- Process automation (Workflows) – you will build a deterministic, reusable workflow (SharePoint + Outlook) that the agent calls as a tool.
- Secure publishing and distribution – you will make the agent available in Microsoft Teams and Microsoft 365 Copilot in line with security and DLP policies.
- Testing – you will use the Preview and Evaluate tabs to debug and measure response quality before deployment.
- Cost and licensing control – you will understand the Copilot Credits model (pay-as-you-go / capacity packs) and estimate consumption before scaling.
In practice, this means that you will leave the course with a ready-to-use, working reasoning agent and the knowledge of how to apply this pattern to real, complex processes in your organization – faster, cheaper, and more securely than building a solution from scratch.
Moduł 1 – Wprowadzenie do agentów i Copilot Studio
- Chatbot vs. asystent AI vs. agent – trzy pojęcia, trzy różne zachowania i wartość biznesowa
- LLM jako „mózg” agenta (tokeny, okno kontekstu, prompting); RAG utrzymujący odpowiedzi w kontekście danych
- Orkiestracja generatywna zamiast drzewa tematów; spektrum autonomii i bezpieczne granice (guardrails)
- Copilot Studio jako platforma; wybór harness: GitHub Copilot vs. Standard vs. Copilot chat
- Bloki budulcowe nowego doświadczenia: Instructions, Knowledge, Skills, Tools, Model, Work IQ, Workflows, Triggers
- Powierzchnia autorska New experience: Build · Preview · Evaluate · Monitor; przygotowanie tenantu i Copilot Credits
Moduł 2 – Rozwiązania Power Platform (Solutions)
- Solutions jako kontenery pakujące agenta i jego komponenty w cyklu ALM
- Rozwiązania zarządzane (managed) vs. niezarządzane (unmanaged) – kiedy używać
- Publisherzy i prefiksy; role bezpieczeństwa (Environment Maker, System Customizer, System Administrator)
- Ład i narzędzia governance: Power Platform Admin Center, Managed Environments, CoE Kit
- Cykl życia rozwiązania DEV → TEST → PROD; Solution Explorer w Copilot Studio
- Laboratorium: utworzenie publishera i własnego, niezarządzanego rozwiązania dla agenta
Moduł 3 – Tworzenie agentów w harness GitHub Copilot
- Lekcja 1 – Tworzenie agentów niestandardowych: opis agenta w języku naturalnym, architektura (Instructions, Knowledge, Skills, Tools, Model)
- Grounding w SharePoint z poszanowaniem uprawnień; potok RAG (query rewriting, retrieval, summarization, walidacja) i moderacja treści
- Laboratorium: budowa agenta AI i ustalenie punktu odniesienia (baseline) w Preview
- Lekcja 2 – Używanie narzędzi (tools): konektory, MCP i przepływy; model sam wybiera narzędzie na podstawie nazwy i opisu
- Laboratorium: dodanie narzędzia SharePoint – Get items i konfiguracja wejść
- Lekcja 3 – Stosowanie umiejętności (skills): wielokrotnego użytku zachowania w plikach SKILL.md (YAML + Markdown) – unikalne dla harness GitHub Copilot
- Laboratorium: dopracowanie umiejętności troubleshooting i device guidance oraz walidacja w Preview
Moduł 4 – Automatyzacja procesów z przepływami (Workflows)
- Nowe doświadczenie Workflows – samodzielne, wielokrotnego użytku i deterministyczne automatyzacje wywoływane przez agenta jako narzędzie
- Klasyczne agent flows vs. nowe Workflows; podział „agent decyduje – workflow wykonuje”
- Anatomia przepływu: trigger, wejścia, akcje konektorów, dynamiczna treść, wyrażenia, wyjścia; 1500+ konektorów, limit 100 s
- Kiedy przepływ, a kiedy pojedyncza akcja konektora; dobre praktyki (małe kroki, czytelne nazwy, wersjonowanie, testy)
- Laboratorium: budowa przepływu obsługi zgłoszenia urządzenia (SharePoint + Outlook) i wywołanie go z agenta; testy ścieżek brzegowych
Moduł 5 – Testowanie, publikowanie i dystrybucja agentów
- Dlaczego testowanie agentów AI jest inne (niedeterminizm, zmienny grounding, wpływ uprawnień); warstwy testów
- Preview (interaktywny debugging, śledzenie planu, testy „braku odpowiedzi”) oraz Evaluate (powtarzalne zestawy testów, metody oceny)
- Publikacja i kanały: Microsoft Teams, Microsoft 365 Copilot, SharePoint, web; publikacja ≠ udostępnianie; republikacja po zmianach
- Pilotaż i zbieranie informacji zwrotnej; monitorowanie i ład po wdrożeniu
- Licencjonowanie i koszty: Copilot Credits, model pay-as-you-go i plany pojemności; szacowanie zużycia (usage estimator)
- Laboratorium: publikacja agenta AI do Teams i Microsoft 365 Copilot oraz test „na żywo”
-
Basic knowledge of Microsoft 365 services (Teams, SharePoint, Outlook) at user level.
-
General understanding of cloud services and of artificial intelligence and large language model (LLM) concepts – an advantage, not a requirement.
-
No programming experience is required – the course follows a low-code / no-code approach.
-
Prior familiarity with Microsoft Power Platform (environments, solutions) is recommended and makes getting started easier.
-
Completion of an introduction to artificial intelligence (e.g., AI-901) or equivalent knowledge is an advantage.
- manual in electronic form available on the platform: https://www.altkomakademia.pl/
- access to Altkom Akademia's student portal
Training method:
-
Lecture: 30%
-
Workshops: 30%
-
Labs:40%
-
Training: English
-
Materials: English