--- title: "Copilot Studio Agents: What You Can (and Can't) Build" description: "Copilot Studio builds AI agents without code. What it can do, where it hits limits, and when you need a coded agent instead. With a real client case." canonical: "https://www.whatsnext-ai.com/blog/copilot-studio-agents" published: "2026-09-17T07:40:18.813Z" updated: "2026-09-30T09:47:56.817Z" --- # Copilot Studio Agents: What You Can (and Can't) Build Copilot Studio builds AI agents without code. What it can do, where it hits limits, and when you need a coded agent instead. With a real client case. ![Microsoft office building lit up at night, with the Microsoft logo on the roof](https://fgzcpjbyiakhjifaciaj.supabase.co/storage/v1/object/public/media/pexels-2498919-4112849.jpg) Copilot Studio is Microsoft's low-code platform for building AI agents: chat assistants that answer questions and run tasks inside your Microsoft 365 environment, without code. Complex logic or custom data processing hits the limits of the canvas. For simple, knowledge-driven work it is fast and effective. It only becomes a problem when your need outgrows what the canvas can handle, and that is exactly what the sales pages leave out. Something stands out about the search results for "copilot studio agents": every page on the first page of results is written by Microsoft or one of its resellers. Not one of them points to Microsoft's own documentation on the limits, even though it states in black and white where the platform stops. This page does: an honest map of what you can build, where the ceiling sits, and when you need a real coded agent. ### What are Copilot Studio agents? A Copilot Studio agent is an AI assistant you assemble with natural language and a visual editor, without programming. You connect a knowledge source, give it instructions, and the agent answers questions or starts a workflow. It runs inside your existing Microsoft environment: SharePoint, Teams, Power Automate. Worth separating: a Copilot agent and a Copilot Studio agent are not the same thing. The full explanation of the Copilot landscape is in [What is Microsoft Copilot?](https://www.whatsnext-ai.com/blog/what-is-microsoft-copilot). If you are still choosing between Copilot and ChatGPT in the first place, that comparison is in [Copilot vs ChatGPT](https://www.whatsnext-ai.com/blog/copilot-vs-chatgpt). In short: Copilot is the assistant you buy, Copilot Studio is the workshop where you build a custom [AI agent](https://www.whatsnext-ai.com/glossary/ai-agent) yourself. Even on forums, practitioners mix the two up, so you are not alone if it feels confusing. ### What can you build with it? This part works well, and it is only fair to lead with that. Copilot Studio is strong for: - FAQ assistants that answer questions from a SharePoint or document knowledge base. - HR and IT self-service: leave policies, password resets, standard procedures. - Assistants connected to Teams, so colleagues get answers without leaving the app. - Simple workflow triggers through Power Automate and its pre-built connectors, for example starting a notification or an approval step. If your need fits within these shapes, Copilot Studio is often the fastest and cheapest route. There is no reason to build anything heavier. ### Where does Copilot Studio hit its limits? This is where it gets interesting, because it is exactly what the sales pages leave out. Microsoft publishes the hard limits itself, on its own [quotas and limits documentation page](https://learn.microsoft.com/en-us/microsoft-copilot-studio/requirements-quotas). A selection: - Instructions for an agent are capped at 8,000 characters. - A maximum of 100 skills per agent. - A maximum of 1,000 topics per agent in Dataverse environments. - Connector payload is capped at 5 MB (450 KB on government GCC plans). - Queries against a SharePoint list only return data from the first 2,048 rows. - Agents cannot write and run code. In Microsoft's own words: "agents can't write and run code." That last one is the core. You can add structured files such as Excel, but answers to analytical questions might, in Microsoft's own words, not be optimal. Translated into practice this means: no custom data logic, no bespoke validation of what comes in, and no audit trail you set up yourself. The moment your process needs more than generating text from a knowledge source, or your data falls outside those limits, the canvas stops. The same applies to traceability. Copilot Studio gives you built-in analytics and conversation logging, but not a full, customizable audit trail of which source fed an answer, which step the agent took and why. For processes where you must be able to demonstrate that, or where you need your own error handling and version control, the built-in features hit their limit. ### When is Copilot Studio enough, and when do you need a coded agent? The call is not "Copilot Studio good or bad", but "does my need fit inside the canvas or outside it". A useful independent check: when we asked ChatGPT what you can and cannot build with Copilot Studio agents, the model itself landed on the conclusion that highly specialised or complex solutions require custom work with specialised tools and experts. That is not our sales pitch; it is where the technology itself points. Stay with Copilot Studio when: the task is knowledge-driven, the data fits within the limits, and you do not need custom logic or an auditable trail. Choose a coded agent when: you need your own data processing or validation, your data volume exceeds the limits, you have to demonstrate a bespoke audit trail, or you want to reason across multiple steps that go beyond a fixed script. ### What building past the canvas delivers An example of where the line sits. Multi Ratio continuously lists second-hand and refurbished products in a Magento catalog of more than 8,000 items. A simple FAQ bot from documentation would fall within the scope of Copilot Studio. But the real need asked for more: a content pipeline that writes the description, USPs, terms, SEO and metadata for each product and readies it for publishing in the webshop. That is custom data logic and integration with an external system, not generating text from a knowledge source, and it does not fit inside a canvas of 100 skills and 8,000 characters of instructions. The result: 14 times faster time-to-online, from two weeks to one day, and one person now owns the whole process instead of three. Read the full [Multi Ratio case](https://www.whatsnext-ai.com/cases/how-multi-ratio-cut-time-to-online-from-two-weeks-to-a-day). That is the shape of a solution that outgrows the low-code canvas. ### Extend a Copilot Studio agent, or have your own built? Start small and honest. If your need fits inside the canvas, use Copilot Studio, it is the cheapest path. If you notice you are hitting the limits, against the 8,000 characters, the 5 MB, the first 2,048 rows, or against can't run code, that is the signal that your process needs a real coded agent. We build those agents in code. That is deliberate: code can be tested, version-controlled and adjusted when your process changes, and you own what runs. We do not finish a stalled Copilot Studio attempt; we build the agent that actually does what you need, from the point where the low-code canvas ends. Not sure which side of the line your need falls on? [Book a free consultation](https://www.whatsnext-ai.com/contact). We will look with you and be honest about whether Copilot Studio is enough or whether custom work pays for itself. If you want a broader view of what we build first, see [our AI services](https://www.whatsnext-ai.com/ai-services) or [having AI agents built](https://www.whatsnext-ai.com/ai-agents). ## Frequently asked questions ### What is the difference between a Copilot agent and a Copilot Studio agent? A Copilot agent is the ready-made assistant you use inside Microsoft 365. A Copilot Studio agent is one you build yourself in Copilot Studio, Microsoft's low-code workshop. The full explanation is in our guide to what Microsoft Copilot is. ### What are the limits of Copilot Studio agents? Microsoft publishes, among others: 8,000 characters of instructions, 100 skills per agent, 1,000 topics per agent, 5 MB connector payload (450 KB on GCC), and SharePoint list queries that only return the first 2,048 rows. The core: agents cannot write and run code. ### Can you build an agent in Copilot Studio that runs code? No. According to Microsoft's own documentation, agents cannot write and run code. For custom data logic, bespoke validation or more complex reasoning, you need a coded agent. ### How much does a Copilot Studio agent cost? Microsoft charges by usage. Copilot Studio is included with Microsoft 365 Copilot, or you buy capacity packs (Copilot Credits) or pay as you go through Azure. Current rates are on [Microsoft's pricing page](https://www.microsoft.com/en-us/microsoft-365-copilot/pricing/copilot-studio). For knowledge-driven tasks within the limits it is a cheap route; once you need a coded agent, the calculation shifts to what that custom work delivers.