Body
Table of Contents
What is an AI agent?
While a traditional AI chatbot acts as a “talker” (a conversational partner which can answer questions), an AI agent is designed to be a "doer." It is a specialized, purpose-built AI that has been given a specialized role and set of instructions, and sometimes specific data to complete tasks with more autonomy.
An agent can (but does not necessarily) possess the ability to perceive its environment, reason through complex goals, make decisions for itself, or take actions on its own to achieve a specific outcome.
Back to top
Agents vs. Chatbots
The shift from "chatbot" to "agent" represents a jump from general conversation to specialized capability.
Feature
|
Traditional Chatbot
|
AI Agent
|
|
Focus
|
General purpose/Q&A
|
Specialized for a specific task or role
|
|
Data Source
|
General web knowledge
|
Grounded in specific files or websites you provide
|
|
Instructions
|
Follows your immediate prompt
|
Follows a permanent "persona" or set of rules
|
|
Interactivity
|
Responds to one-off questions
|
Can work through multi-step processes
|
| Autonomy |
None |
Can range from none to full autonomy |
Back to top
Dangers and Risks of Agentic AI
Because agents can be given the authority to process data, reason across multi-step workflows, or act as a representatives, they introduce unique risks:
-
Chained Errors: Because agents often follow a multi-step plan, a small misunderstanding in initial steps can "cascade," leading to a significantly flawed final result that might look correct at first glance.
-
Over-Reliance: It can be easy to stop critically evaluating an agent’s output if the agent is perceived as a "subject matter expert," leading to a decrease in critical human oversight.
-
Data Integrity: If an agent is grounded in an outdated version of a policy or syllabus, it will confidently provide incorrect information.
-
Algorithmic Bias: Agents inherit the biases of their underlying models. If used for tasks like initial resume screening or drafting student feedback, they may replicate unfair patterns. As always, human review is crucial.
A note on autonomy: The agents currently available when using Copilot with your Stonehill account are not autonomous, meaning they cannot take actions on their own, only respond to your prompts. This is an important distinction because the term "agentic" overlaps heavily with "autonomous", but not completely.
This distinction is a subtle, but critically important one. Autonomous agents present myriad security and privacy issues, as they are able to take actions in the real world, independently and without human supervision. For this reason, you should exercise extreme caution when considering using third-party agentic AI tools and software.
Currently, the autonomous agentic capabilities of Copilot are being evaluated by the Department of Information Technology and are not available when using Copilot with your Stonehill account.
Remember: Stonehill students and employees should never use third-party AI tools on Stonehill computers or with College data.
Back to top
Using Agents in Copilot with your Stonehill account
All Stonehill faculty, staff, and students have access to basic (non-autonomous) agents through Copilot via their Stonehill account. These tools include the same Enterprise Data Protection that secures your OneDrive and SharePoint data here at Stonehill, meaning your Copilot activity remains within Stonehill’s secured M365 environment, and your chats are never used to train AI models.
Out-of-the-box agents
Microsoft provides several pre-built agents found in the "Agents" tab or via the side panel, including but not limited to:
- Writing Coach: Refines drafts for tone and clarity.
- Idea Coach: Brainstorms agendas and project plans.
- Learning Coach: Provides personalized study tips and practice exercises (specifically designed for students).
Custom agents
You can build a custom agent for your specific needs (e.g., a "Department FAQ agent", or a “syllabus formatter”, or a “test prep tutor”) without any coding:
-
Open the Agent Builder: Click "Create an agent" in the Copilot side panel.
-
Give Instructions: Tell the builder what you’d like your agent’s role to be, and give it any starting instructions you’d like it to operate from in each interaction. For example: "You are a lab safety assistant for the Biology department. Only use the information in the provided safety manual to answer questions."
-
Add Knowledge: Upload the specific files or links to publicly-accessible webpages you want the agent to reference in its operations.
-
Test and Share: Use the preview window to test your agent out. Ensure it follows your rules to your satisfaction before sharing any links to it with others.
Back to top