What are the benefits of Copilot Agents?
You could just use the general version of Copilot to do many of the same things that you can do with Copilot Agents. The main benefits of using Agents are:
- For repetitive tasks, or tasks shared between several people in a team, you can store the task instructions in the Copilot Agent, so there’s no need to keep writing these out each time.
- If everyone in a team is using the same Copilot Agent for the same task, it will provide a more standardised and consistent output than if everyone using their own private AI chats – especially if you carefully structure the instructions, for example by specifying output format, or providing content that it must use in its answers.
- You can give Copilot Agents ‘Starter prompts’ – these are essentially ‘buttons’ for conversation starters, so you can guide the user as to what functions the Agent has. With general AI chat tools, the user has to start with a blank text box, so they need to have some idea of what instructions to give it.
- One great use of AI is for role-plays. With Copilot Agents, you can hide the description of the role-play and characters from the user, which helps with suspension of disbelief and immersion.
- Another good use of Copilot Agents is to create student-facing activities – you can add appropriate-level, accurate information for it to use in its feedback, rather than relying on AI’s ‘general knowledge’.
- You can give Copilot Agents access to specific websites (these must be ‘open’ – i.e. no log-in required).
What are the limitations of Copilot Agents?
Coventry University staff / students have access to the version of Copilot that is included in the ‘Education Microsoft 365 licence’. This version of Copilot allows us to create and use ‘Agents’ that are similar to other custom bots, such as OpenAI’s ‘custom GPTs’ or Google Gemini’s ‘Gems’, but with some important limitations:
- You can’t give Agents access to files. You need to put all the custom information you want the Agent to have access to – i.e. information that it won’t have from its general training data – in the ‘prompt’ (which can be up to 8,000 characters), or give it access to specific websites.
- You can give Agents access to general ‘Web Search‘ or specific URLs, but they can only access ‘open’ websites (i.e. websites that don’t require a log-in).
- You can’t connect Copilot Agents to other apps.
- You can’t share Copilot Agents externally – they can only be shared with Coventry staff/ students.
- Copilot Agents can’t currently create images or data visualizations (the normal web version of Copilot can do both of these things, so if you need that functionality you’d be better off sharing a prompt than creating a Copilot Agent).
- Copilot Agents are tied to the Microsoft account of the person who made them, so if that person leaves the organisation, the Agents will stop working. I recommend creating an index with all the information needed for someone else to recreate any Copilot Agent that you make.
- In order to set expectations about Copilot Agents more generally, it’s useful to note that Microsoft’s use of the word ‘to ‘Agent’ to describe their custom bots is a bit of an outlier. In general, ‘Agent’ or ‘Agentic’ is used for AI tools that can operate with a degree of autonomy – for example ‘narrow’ agents like Google’s Deep Research that can create research reports, or ‘general’ agents like ChatGPT’s Operator that can use a computer. Copilot’s Agents are not ‘Agentic’ (autonomous) in this sense. It’s best to think of Copilot Agents as standard chat bots with custom instructions.
- For more general limitations of Copilot vs other AI tools, such as restrictions on the amount of data/ number of words it can work with, see the FAQ.