For the complete documentation index, see llms.txt. This page is also available as Markdown.

What's Copilot Studio?

Session 1.1 — what the platform actually is, the three harnesses underneath it, and how they map to the knowledge-vs-skill frame from "What Is AI, Really?"

Group 1 · Session 1.1 of the Copilot Studio course. Full plan on the Mission page.

Starting from my own frame

Your piece on AI put it plainly: knowledge by itself doesn't make something human-like, it makes it a library. Skill is what sits between wanting something and having it, and skill only gets built through repetition, failure, and watching whether the thing you made actually holds up.

That distinction turns out to be the most useful lens for this whole course, because Copilot Studio is not one tool that does one thing. It's a studio that can build a library — an agent that answers questions accurately from documents — or it can build something closer to skill: an agent that takes a goal, works through the steps, hits a wall, and tries something else. Which one you get depends on a choice you make early, and most tutorials skip past it.

The plain definition

Copilot Studio is a low-code studio — a single web app at copilotstudio.microsoft.com — for building and managing AI-powered agents and workflows, connecting them to your organization's data and systems, and publishing them to the channels people already use: Teams, websites, Microsoft 365 Copilot, mobile apps.

Inside it you can build three kinds of things, and they can call each other:

Building block
What it is

Agent

An AI assistant that handles a conversation and completes tasks. It follows the instructions you give it, draws on the knowledge you connect, and uses tools to act — reasoning about what the user needs and deciding the next step.

Workflow

A drag-and-drop automation. Each step can reason and act, mixing agents, tools, and plain logic in one flow — the predictability of automation with some AI flexibility layered in.

Agent flow

Copilot Studio's native version of a Power Automate flow. Runs standalone, or gets attached to an agent as a tool the agent can call and get a result back from.

The part that actually decides how your agent behaves

Everything you build runs on what Microsoft calls a harness — the runtime sitting between your design and the underlying model. It decides when to call the model, what to hand it, how to read what comes back, and which tools to trigger. The harness you pick changes how your agent reasons, how much it can do in one turn, what it can do out of the box, and how it's billed. There are three.

Harness
Best for
How it works
Maps to

Standard

Rule-based, structured conversations

Matches requests to the topics you define; falls back to generative answers from connected knowledge

Mostly the library: retrieves and answers

GitHub Copilot

Reasoning-heavy, multistep processes

Breaks a goal into steps, calls tools across connectors and other agents, adjusts when a step fails instead of stalling

The doing: iterates, recovers, adapts

Copilot chat

Extending Microsoft 365 Copilot Chat

Grounds M365 Copilot Chat answers in your enterprise knowledge, published internally only

A library card scoped to one building

Notice the middle one is the only harness Microsoft itself describes with the phrase "agentic reasoning." That's not a coincidence — it's the harness built around your second question, the one about doing rather than knowing.

Worked example: the same request, two harnesses

Your course scenario is Northwind Outfitters, a retailer whose agent handles order status, FAQs, and returns. Here's the same customer, two different requests, and why they land on different harnesses.

Library question — standard harness

"What's your return policy for worn shoes?" This matches a topic or gets answered straight from a connected FAQ document. No steps, no judgment calls, the same correct answer every time. Sessions 2 through 4 of this course build exactly this.

Doing question — GitHub Copilot harness

"My order never arrived. Can you check what happened, and if it's lost, start a replacement and email me once it ships?" That's a lookup, a branch, an action, and a follow-through — four steps where the third depends on what the first one finds. This is what Group 5 and 6 build toward.

Check your retrieval

Answer before you look anything up. Getting one wrong is more useful right now than getting it right.

An internal IT help-desk agent that must give the same correct answer to "how do I reset my VPN password" every single time, with zero surprises.

  • GitHub Copilot harness

  • Standard harness

  • Copilot chat harness

Standard harness

An agent that reads incoming invoices, matches each one to a purchase order, and routes anything that doesn't match to the right approver — adjusting when a match is ambiguous.

  • GitHub Copilot harness

  • Standard harness

  • Copilot chat harness

GitHub Copilot harness

An agent that only answers employee questions from SharePoint policy documents, reachable exclusively inside Microsoft 365 Copilot Chat.

  • GitHub Copilot harness

  • Standard harness

  • Copilot chat harness

Copilot chat harness

True or false: every agent in Copilot Studio uses the same reasoning engine underneath, no matter what you build.

  • True

  • False - the harness you choose is that engine

False - the harness you choose is that engine

Reflection

In your own knowledge-versus-skill frame: which harness is the library, which is the skill - and where does the Copilot chat harness actually sit? It doesn't map cleanly to either.

ONE REASONABLE ANSWER Standard harness is the library - it retrieves and answers, reliably, without independent judgment. GitHub Copilot harness is the skill - it takes a goal, tries steps, and adapts when one fails, which is the repetition-and-failure loop your essay describes. Copilot chat harness doesn't sit between them; it's arguably an even purer library than the standard harness, because Microsoft built it to optimize for one job only - connecting a person to information inside Microsoft 365 Copilot Chat - with no ambition to complete multistep tasks at all.

The single best next read is the primary source this lesson is built on: Microsoft's own Choose a harness guide. It has the full comparison table — billing, file handling, error recovery, publishing — for all three harnesses side by side.

Sources

Last updated