⚙️ Lab note (added September 2026): every agent in this series was built and run in the new Copilot Studio experience — what Microsoft now calls the GitHub Copilot harness (GA August 2026). The classic “standard harness” is a different build experience with different behaviour — my standard-harness rebuild and full harness comparison are scheduled on this blog for late October 2026.
Every second conversation I have these days starts the same way: “we want an agent, where do we even begin?” The good news is that you no longer need a development team or a six-week program to find out. I recently sat down and built a working agent from scratch in the new Copilot Studio experience — connected to my own environment, grounded on real content — in a single evening, on trial licenses. This post walks through the first hour of that build: creating the agent, giving it instructions, and teaching it which websites it is allowed to learn from.
A word of warning before we start: if you have read Microsoft’s documentation or older blog posts, unlearn the screenshots. The new Copilot Studio experience looks nothing like the classic topic-and-trigger designer. There are no topic canvases on day one — you get an instructions page, a model picker, and a right-hand rail for channels, tools, and knowledge. It feels less like building a chatbot and more like briefing a new hire.
⚠️ Important: the new experience runs on Copilot Credits, and since September 1, 2026 that is enforced — my environment hit a hard “out of credits” wall mid-build. Licensing and trials deserve their own walkthrough, and that is exactly what next Tuesday’s post covers. For today, assume your tenant has Copilot Studio trial capacity in place.
Step 1 — Create the Agent in the Right Environment
Head to copilotstudio.microsoft.com and check the environment picker in the bottom-left corner before you click anything else. Copilot Studio drops you into your default environment, and agents — like flows and apps — live where you create them. I switched to my test sandbox first. One quirk worth knowing: the picker searches display names, not URLs, so if your environment’s friendly name doesn’t match its address (mine didn’t), verify with a quick WhoAmI call against the org URL.
From Home, select the Agent tile. You land directly in the editor with an untitled agent and a blinking name field. No wizard, no template gallery — the product wants a name and instructions, in that order.
Step 2 — Name It and Write Instructions Like a Job Description
I named mine Blog Lab Agent and treated the Instructions canvas as a job description: who the agent is, what systems it may touch, when it should ask instead of act, and what it must never invent. Mine ended with a line I recommend stealing: “If information is missing, say so and ask one clarifying question. Never invent SKUs, prices, or customer data.” Instructions are the cheapest guardrail you will ever deploy.
Notice the Model dropdown at the top of the rail. My tenant defaulted to Claude Opus 5 — yes, Copilot Studio now runs frontier models from more than one vendor, and you can swap them per agent. The days of “one model fits all” are over.
Step 3 — Add Website Knowledge (and the Toggle That Blocks Everyone)
Knowledge is what the agent is allowed to read. Click + next to Knowledge and you get the featured sources: public websites, SharePoint, OneDrive for Business, and friends.
Select Public websites, and here is the quirk that will cost you five confused minutes: the dialog opens with Search all websites turned on, and while that toggle is on, the URL field silently refuses your input. Turn it off first, then add your sites. I added learn.microsoft.com/en-us/dynamics365/ and my own blog — each becomes a named source with an editable description that tells the agent what it is for.
Key insight: website knowledge is not a “connection”. There is nothing to authenticate, no actions to call — the agent simply runs a web search scoped to the domains you listed and cites what it finds. Read-only, public content, zero credentials. That distinction matters a great deal once tools enter the picture, and it is the subject of next week’s deep dive.
Step 4 — Save and Say Hello
Hit Save (Ctrl+S works), switch to the Preview tab, and greet your agent. At this stage it can hold a conversation, answer from its instructions, and quote your chosen websites with citations. It cannot touch your data yet — that is deliberate, and that is where MCP tools come in later in this series.
What I Learned
Building the shell of an agent is now genuinely a fifteen-minute job — the real work has moved from “how do I wire this up” to “what should its instructions say and what should it be allowed to read.” Write the job description carefully, scope the reading list deliberately, and you have a foundation worth building on. An agent with good instructions and narrow knowledge beats a genius with the whole internet.
Next in this series: the licensing story — trials, the September credit wall, and who needs which license. Bring your billing admin.
