Agents are a type of AI that, once given a goal, acts autonomously to complete that goal, without being told which steps to take, or how to try a different route if the first attempt does not work out.
Agents are not websites where you ask a question and get an answer. An agent is a type of AI that actually DOES things for you.
Different Definitions
AI companies and media pundits are using the term “agent” to mean different things:
- Copilot has a button to create agents, but these are little more than pre-programmed ground rules for how it reacts to prompts. For instance, if you set one up and instruct Copilot to always answer in pirate-speak, when you prompt Copilot through that agent, you don’t need to mention pirates to get the outcome of what you asked for in pirate-speak.
- A more robust application of AI can be thought of as automated workflows. One example might be a company looking to automate all the steps of Accounts Payable, where a human doesn’t need to manually perform each step. The process runs by itself, but very often the steps to take are spelled out ahead of time, so the software isn’t truly autonomously figuring out how to accomplish the goal. The term “agent” is being applied because it’s a hot buzzword.
- Truly agentic AI can appear as an agentic browser or an agentic desktop app.
- An agentic browser will, once given a goal, carry out the tasks without being told which programs to open or how to interpret anything. One example might be asking an agentic browser to find the cheapest tickets to a local concert venue you specify; it will move the cursor, open websites, and type things into online forms without the human user touching the mouse or keyboard. If given the right information, an agentic browser can complete online assignments and discussion board posts.
- An agentic desktop app similarly can open files and programs on your computer’s hard drive, making its own decisions as it goes, all while trying to complete the overall goal it was given.
Data Privacy and Security Concerns
Agents are often a complete wildcard when it comes to data privacy. Many will transmit information back to its company, as a way to continually refine and improve the model. Some of them will also sell your information to third parties such as advertisers.
The problem is, for an agent to do useful things, you often need to give it your credentials. It will learn your email username and password if you ask it to sort, summarize, or answer emails. If you ask it to take a test inside Webcourses for you, you have to give it your NID and NID password. If those are sent back to the company or sold to third parties, now your credentials offer them a way to access UCF data. Not all agentic AI is trustworthy. Some have gaping security problems. A few may even have malicious intent.
Hallucinations and Misunderstandings
Agentic AI still relies to some extent on LLMs, and thus probabilistic training. It predicts one step at a time, and predictions can be incorrect. Hallucinations in a block of text or in an image are one thing, but hallucinations can be much more serious when setting an agenda loose to do things in your browser or on your own computer files. If it deletes a wrong file, you might never know about it.
Other misunderstandings can occur in the quest for efficiency and optimization. Most AI is designed to use the fewest tokens and resources possible in order to complete the task. Sometimes that means invisibly reducing a large task to a simpler plan. Important directions might get lost in that optimization. In 2026, one security engineer directed an agent to delete certain outdated emails but confirm with the human user before each deletion, yet that confirmation directive was lost as the AI optimized the task, and just deleted all emails it could find systematically. The engineer had to unplug to computer to stop the agent.
Bottom Line
Today, there are no safe examples or applications of agentic browsers or agentic desktop apps. UCF students, faculty, and staff are urged to avoid installing or using agents on UCF devices or any device that connects to UCF data.