From Chatbots to Agents: What School Psychologists Need to Know
Key Takeaways
Agents have “arms”: approved tools that let them act on a computer rather than only reply to questions.
A chatbot usually waits for a person to begin and continue the conversation. An agent may be scheduled, triggered by an event, or asked to repeat an assignment.
Those arms can create documents, research on the web, organize files, use applications, and—when specifically permitted—operate a computer in many of the ways a person can.
An agent’s ability to act is both its primary benefit and its primary source of risk.
School psychologists should begin with low-risk uses involving public, fictional, synthetic, or otherwise nonsensitive information.
When I talk with school psychologists about artificial intelligence, most are still using it primarily as a chatbot. They ask questions, generate drafts, summarize information, or brainstorm ideas. Those uses can be valuable, but they do not fully represent where AI is now—or where many experienced users and AI enthusiasts are already heading.
I include myself in that group of users who are moving beyond chat alone. I am not claiming to be an AI expert. I am a school psychologist, researcher, professor, and consultant who has been experimenting with these tools, paying attention to what they can do, and thinking carefully about what responsible use should look like in our profession.
This post begins a series on AI agents. The series will examine what agents are, how they differ from chatbots, how tools and permissions affect what they can do, how goals and repeated steps shape their assignments, and what I have learned from building a research-to-practice agent. I think these developments are important for school psychologists who want to understand where AI use is going, even if they are not ready to build or use an agent themselves.
This is not the first time I have written about agents. In October 2024, I posted The Future of AI Agents in School Psychology. At that point, I described agents largely as a future possibility and imagined what it might look like for a school psychologist to use several of them five years later.
The transition arrived faster than I expected. By late 2025 and early 2026, agents were moving from something I discussed as a future development to something I was using in practice. I have been using agents since at least February 2026, and at NASP 2026 I found myself talking with several people who work closely with AI about how I was already using them. What felt speculative in 2024 now feels much more immediate.
This series actually began while I was out taking a walk. I talked through the ideas with Codex using voice, questioned suggestions, changed direction, and gradually shaped the project without sitting in front of a computer. I was getting exercise, enjoying the conversation, and developing material I could later turn into blog posts, training resources, and practical tools. It was productive, but it was also fun.
What made the experience especially useful was not voice alone. With my permission, Codex could consult relevant project files and use authorized context from earlier conversations. It could also create and revise documents rather than merely suggest what I might write. That combination of conversation, context, and action captures part of what is changing about how we can work with AI.
This is a screenshot of Codex helping me create the Substack version of this post. It is a small but concrete example of the distinction I am describing: the AI was not merely suggesting text. With my permission, it was working inside the application, formatting the post, adding the image, and saving the result as a draft for my review.
Agents Are AI with Arms
I think of an agent as AI with arms. A chatbot primarily talks back to you. An agent can be given tools that allow it to take action: researching on the web, creating a Word document, organizing files, using applications, entering information into a website, or—if a person grants broad permission—operating much of a computer in ways that resemble what a person can do.
The arms are not unlimited by default. Their reach depends on which tools and permissions a person provides. That is the benefit and the risk: an agent can do more useful work, but a mistake can also have consequences beyond an inaccurate answer.
Instead of asking AI for one response, we can give an agent an assignment and allow it to work through a series of steps. Depending on how it is configured, an agent may search approved sources, organize files, run code, check its own work, update a dashboard, or prepare something for human review.
That shift—from answering to acting—is what makes agents useful. It is also what makes them more consequential.
What Is an AI Agent?
OpenAI offers a straightforward definition: agents are systems that independently accomplish tasks on a user’s behalf. The important word is not independently in the sense of having unlimited autonomy. The important idea is that the system can manage parts of a multistep assignment rather than waiting for a person to direct every individual step.
Imagine that you want to stay current on new research related to artificial intelligence and school psychology.
With a chatbot, you might paste an article into a conversation and ask for a summary. That can be helpful, but you still have to find the article, decide whether it is relevant, organize it, and remember to return to it.
An agent could handle a clearly limited assignment:
Check a defined set of public research sources.
Remove duplicate articles.
Sort the remaining articles by relevance.
Prepare short, practice-focused summaries.
Identify the country and sample represented in each study.
Add the articles to a searchable dashboard.
Present a small number for review rather than delivering a daily fire hose.
That is more than a response. It is a job with a sequence of steps, rules, and a stopping point.
A Chatbot Answers; an Agent Pursues an Outcome
The distinction is not perfectly clean. Some chatbot products now include scheduled features, and not every agent is configured to run automatically. Still, this is a useful starting point:
Here are eight practical differences:
Chatbot: Responds to a prompt. Agent: Works toward a stated outcome.
Chatbot: Usually begins when a person opens the chat and sends a prompt. Agent: May begin on a schedule or when a specified event occurs.
Chatbot: Usually completes one conversational request. Agent: May take multiple steps and use multiple tools.
Chatbot: Waits for the next instruction. Agent: Can select an appropriate next step within its established limits.
Chatbot: Requires the user to return and prompt it for each new request. Agent: May repeat an approved assignment without being prompted each time.
Chatbot: Produces an answer or draft. Agent: May create, organize, check, or update work products.
Chatbot: Primarily interacts through a conversation. Agent: May act in the background and return with results or a request for approval.
Chatbot: Usually relies on the current conversation. Agent: May use durable instructions, saved state, and an approved working space.
When an agent begins on a schedule or after an event, it is not spontaneously deciding to wake up. A person has previously established the schedule or triggering event, the assignment, the tools it may use, and the points where it must stop for approval.
Regular chat is often the better choice for brainstorming, reflection, or a one-time question. An agent becomes more useful when an assignment is repeatable, involves several steps, and is structured enough that you can define what success looks like.
Not every task needs an agent. Sometimes a conversation is exactly what we need.
The AI Intern Now Has Tools
I have previously described AI as being like an intern: fast, useful, uneven, and always in need of supervision. That analogy becomes even more important with agents.
A chatbot is a bit like asking an intern a question at your desk.
An agent is more like giving that intern an assignment, access to selected tools, a working space, a deadline, and instructions about when to return for approval.
The intern may now be able to open files, organize information, run analyses, or prepare communications. That added capability can save time—but only when the assignment and authority are clear.
You would not give a new intern unrestricted access to every student record, permission to email anyone, and authority to delete files. You would identify the task, limit access to what is needed, review the work, and expand responsibility only after the intern demonstrates reliability.
Agents should be treated the same way.
For more on this analogy, see AI Is Like an Intern: Useful, Uneven, and in Need of Supervision.
What Could a School Psychologist Use an Agent For?
The safest starting points do not involve student information. Examples include:
A research-to-practice agent that monitors public research, prioritizes relevant articles, and prepares concise summaries.
A professional-development builder that turns selected public research into draft slides, speaker notes, handouts, and activities, then reviews nonsensitive participant feedback and prepares proposed improvements for the next presentation.
A policy and guidance monitor that tracks public updates from state agencies and professional associations, summarizes meaningful changes, and updates a dashboard for review.
A resource-library curator that organizes public intervention and family resources by topic, audience, or evidence level. In a future post, I plan to explore how Sites might turn a resource library like this into a simple public website for parents and caregivers.
A consultation-practice simulator that uses entirely fictional scenarios to help graduate students practice reflection, empathy, and collaborative problem solving.
These examples vary in complexity, but they have something important in common: the assignment can be clearly defined, the information can be limited, and the work can be reviewed before it affects another person.
If you want to see another concrete example of AI acting rather than merely answering, see AI for Program Evaluation: How School Psychologists Can Build Better Surveys Faster. In that example, Codex helped configure survey response requirements, display and skip logic, carry-forward choices, validation rules, and testing—not simply draft survey questions.
Start Small
My recommendation is to begin with one task that is annoying, repeatable, and low risk.
Do not start by asking, “How can I automate my entire job?” Instead ask:
What do I repeatedly spend time finding, sorting, or reformatting?
Can the task be completed without confidential information?
Can I describe what a good result looks like?
Can I review the result before anything consequential happens?
Would an agent meaningfully improve this over a normal chat?
The next post will address the safeguards surrounding an agent: its tools, permissions, separate working folder, approval points, and limits. Later posts will examine goals and loops and provide a closer look at the research-to-practice agent I have been developing.
I am also curious whether school psychologists are paying attention to this shift. Are you already experimenting with agents, interested but cautious, or unconvinced that they would be useful in your work? Your response will help me understand how much interest there is in more practical training and examples.
Examples of AI Agent Platforms
For the purposes of this series, I will focus primarily on ChatGPT Work and Claude Cowork. These are the two platforms most relevant to the kinds of computer, document, research, and application-based assignments I will be describing.
They are not the only options. Other examples include:
Gemini Spark, Google’s personal AI agent for ongoing and scheduled tasks.
Manus, a general-purpose agent that works in a cloud computer and can also interact with local computers.
Perplexity, an AI answer engine that has expanded into agent capabilities through Perplexity Computer and the Comet browser.
Genspark Super Agent, a general-purpose agent from a company that initially launched as an AI search engine.
The features, access requirements, and limitations of these products change quickly. Including them here is meant to provide concrete examples, not to endorse every platform or suggest that each is appropriate for professional or confidential information.
Final Takeaways
Think of an agent as AI with arms: it can use approved tools to act, not merely respond.
Agents may be scheduled, triggered by an event, or asked to repeat an assignment, but only when a person has configured those capabilities.
Use an agent when the assignment is repeatable, involves several steps, and has a result that a person can review.
Begin with public, fictional, synthetic, or otherwise nonsensitive information.
Remember that greater ability to act creates both greater usefulness and greater risk.
A Brief Caveat About Usage Limits
You do not need advanced technical skills or an expensive subscription to begin experimenting with agents. ChatGPT Work and Codex are currently available through ChatGPT’s free and paid plans, including the $20-per-month Plus plan. Availability may differ in school-managed accounts. One practical limitation is that longer, more complicated agent tasks can use your plan’s available usage more quickly than an ordinary chat.
AI Disclosure: Generative AI was used to assist with drafting and editing this post and to create the accompanying image. I reviewed and revised the content and take responsibility for its accuracy and final form.