AI Agents

[argosanti]

I was recently approached by a colleague in school who was wondering if we had ever thought about providing tools to be able to create AI Agents to support repetitive activity. I've heard of the concept but haven't seen any examples of tools or how they're used in Education. Does anyone have any good examples or insight?

[Andycat]

Not personally, but Hannah Fry has just done a great YT video on them!

[TheHyperTechie]

Gemini Gems could do this.

I have tested/created some myself and they work well. 

[argosanti]

29 minutes ago, Andycat said:

Not personally, but Hannah Fry has just done a great YT video on them!

 

Thank you! I will take a look.

[argosanti]

15 minutes ago, TheHyperTechie said:

Gemini Gems could do this.

I have tested/created some myself and they work well. 

 

What kind of use cases have you been using them for? Is it pricey? 

[synaesthesia]

Would it be a new fangled, more expensive version of creating FAQs, knowledgebases, wikis, slacks etc?

I.e. great intention but has next to zero impact because a) noone has time to pay attention to them and b) humans like a personal touch? From another human of course, rather than a computer trying to behave like one.

 

/me mutters something about young whippersnappers, this was all fields and freddos were only 10p...,

Edited May 7 by synaesthesia

[argosanti]

1 minute ago, synaesthesia said:

Would it be a new fangled, more expensive version of creating FAQs, knowledgebases, wikis, slacks etc?

I.e. great intention but has next to zero impact because a) noone has time to pay attention to them and b) humans like a personal touch? From another human of course, rather than a computer trying to behave like one.

 

/me mutters something about young whippersnappers, this was all fields and freddos were only 10p...,

 

The examples given to me where things like signposting people to the correct information e.g. "i'm having a baby" might refer someone to the relevant maternity docs with the process explained;  and the other was around constructing social media wording for job packs in the correct tone of voice etc.

 

I'm sure they'll be other use cases, we do use Microsoft Flow for limited automations internally but seems to be a buzz word at present that AI Agents can save the company loads of time but not seeing many live examples.

[TheHyperTechie]

7 minutes ago, argosanti said:

 

The examples given to me where things like signposting people to the correct information e.g. "i'm having a baby" might refer someone to the relevant maternity docs with the process explained;  and the other was around constructing social media wording for job packs in the correct tone of voice etc.

 

I'm sure they'll be other use cases, we do use Microsoft Flow for limited automations internally but seems to be a buzz word at present that AI Agents can save the company loads of time but not seeing many live examples.

Well... we are on Google Workspace Education Fundamentals, which is free.

 

You could definitely build a Policy Navigator with this though - you essentially attach the policies and the AI bot uses this as knowledge to reply to the user e.g. "I am ill ..... > Please refer to our sickness policy here". 

[argosanti]

Just now, TheHyperTechie said:

Well... we are on Google Workspace Education Fundamentals, which is free.

 

You could definitely build a Policy Navigator with this though - you essentially attach the policies and the AI bot uses this as knowledge to reply to the user e.g. "I am ill ..... > Please refer to our sickness policy here". 

 

We are hybrid at present - Microsoft for everything and Google Workspace for Flex/Chromebooks. Interesting though - is there a cost? Microsoft offers some Copilot functionality but costs do ramp up for something similar?

[TechMonkey]

In M365 there is a whole agents section, which I have been assured others schools are using really affectively for things, but I can't get any info about licensing needs, decent documentation of how to use them, or any clear idea from the staff suggesting them of what 'things' are wanted.

 

It is frustrating as I'm sure they may be useful but there seems to be so much fud around with equally as much lack of solid information.

[TheHyperTechie]

Just now, argosanti said:

 

We are hybrid at present - Microsoft for everything and Google Workspace for Flex/Chromebooks. Interesting though - is there a cost? Microsoft offers some Copilot functionality but costs do ramp up for something similar?

No cost, as far as I am aware (it hasn't cost us anything!).

I can't comment too much on anything Copilot related to be honest, as most of my AI experience is with Gemini - as we are a Google School.

[Pottsey]

1 hour ago, argosanti said:

 

The examples given to me where things like signposting people to the correct information e.g. "i'm having a baby" might refer someone to the relevant maternity docs with the process explained;  and the other was around constructing social media wording for job packs in the correct tone of voice etc.

 

I'm sure they'll be other use cases, we do use Microsoft Flow for limited automations internally but seems to be a buzz word at present that AI Agents can save the company loads of time but not seeing many live examples.

My experience yes it saves the company loads of time by wasting the customers time and driving up customer frustration levels. Almost every time I have encountered an AI agent its made the overall experience worse.

[E_G_R2]

I used copilot to make an agent to answer questions about our school ONLY from the school website.  Couldn't even get the head teachers name right!

[dhicks]

4 hours ago, argosanti said:

I've heard of the concept but haven't seen any examples of tools or how they're used in Education.

 

The more recent new items have been about things like OpenClaw, a "personal AI assistant". It's software that runs on a device somewhere (your local machine or a server - seemingly the Mac Mini works well) and can use an LLM (one of the public cloud ones, or possibly a local one if you have the hardware available to run one) to do whatever tasks you can think of to give it. You can add "skills" and, via MCP servers, give it ways of interacting with external systems (websites with APIs, I think sometimes vision-related capabilities, etc). The more ambitious projects have done things like handing it access to a bank account and email server, pointing it at the internet and have it run a coffee shop. This kind of thing is very much in its early days at the moment - really just since teh start of this year. Most of the stories about interesting / succesful projects seem to have been by developers working for larger organisations, where spending £20,000 on LLM server (token) costs in a month is seen as an interesting side project.

 

In our school setup, we have a Google Workspace domain, with Gemini licensing for most staff. At the moment, we have a number of staff using Gemini features in email and in Google Slides to help produce presentations ("Computer, turn this document into a slideshow, with some illustrations along the way"), which seems to be working quite well. We've had a few people try out NotebookLM, which seems like it should be a great tool for exam revision ("Computer, take this pile of lessons notes and process them into an interactive document I can ask questions of"). I've tried installing the Gemini command-line assistant in a development environment, but that seems to require a different billing setup for some reason, so I haven't got much further with that yet. I find Gemini useful for answering detailed programming questions ("Computer, how do I go about writing a custom Java authentication plugin for the Guacamole remote desktop server?"), although I don't have much else to compare it against as yet. I aim to try, as someone has suggested, the Gemini "programming assistant" Gem when I can get a suitible environment set up.

[jmak]

Your original post mentions crafting your own AI agent. That is achievable but non-trivial and introduces significant risk. The Hannah Fry YouTube video is a very good demonstration of that. If you mean using a commercial agentic AI product on your environment, then that's an entirely different prospect - but still significant and potentially consequential.

 

I don't work in a school any more, but our team is making extensive use of Claude - often for "traditional" AI tasks like analysing and summarising documents, producing user guides and knowledgeable articles. The recent change for me is using the agentic features. Yesterday I used the Claude browser plugin running on a tab where I was signed in to Jamf with a prompt along the lines of "create a configuration profile which sets an iPad to run in single app mode to launch x webpage and some it to this group of iPads". It then worked it's easy round the Jamf console clicking buttons and produced the profile, applied it to a group, pushed it out and the iPad on my desk restarted and launched the webpage. I did similar with Jira last week: create a form to gather these details and apply it to a new request type called "request service a" and add it to the portal in x space. That needed a bit more tweaking and I wouldn't have released that to general availability before proper testing, but it did work.

 

I'm finding it genuinely useful, but one of our key purposes in using it is to try to get ahead of our users to find out what gates and guardrails we might need to put in place. Web based learning just became (even more) pointless. As an experiment I opened one of our learning packages in a browser and gave the prompt "watch the videos and complete the course. Then take the quiz". Ten minutes later a certificate of completion popped into my (and my manager's) inbox without me viewing any of the content or any further interaction. 

 

I mentioned the browser extension - that's because I run an ARM based windows machine so the Claude Cowork application won't install. On other Windows clients, MacBooks and Linux clients the installed version will operate on any application or setting you ask it to, so much more powerful - I'm thinking "add a registry key" or "create a PowerShell script to do x and run it". Quite a step on from "Gemini, please tell me the steps to edit the registry" or "create a PowerShell script to do x". You may have your environments locked down, but most users who'd asked an AI how to achieve something would get a fairly good set of instructions on the standard way to do things which you've probably got blocked. With agentic AI, it'll just keep searching until it finds a way through, which is a different risk profile.

 

We've set up an autonomous AI agent in our help desk. It only runs out of hours and one of the first tasks in the morning is to review the tickets it's answered - not because it gets a lot wrong, but because we (currently) want our users to feel they're getting a personal service. It looks at our knowledgeable and previous ticket responses to attempt to answer the queries and the agent got its first five star review last month. Some users will hate it, others value a solution to their problem out of hours - the alternative is no answer at all until the help desk opens.

 

Our team's paid subscription to Claude is the only group pro licence registered. Anyone else signing up to Claude as an individual with their work domain email address gets a prompt asking them if they want to join our team. We've received hundreds of requests to join in the last month - to me that indicates there's a real appetite for it and that we need to be ready for our users to have access to it whether we want it or not. All they have to do is sign up to a service with a personal email address and then they can run it in a browser to access any of your systems they have a login for.

[dhicks]

7 hours ago, jmak said:

We've set up an autonomous AI agent in our help desk.

 

Interesting - which helpdesk are you using? We are currently using Trello as a (reasonably basic) helpdesk, which currently seems to fit the way we work quite well - we have it hooked up to emails to people can email tickets in and we can reply, and I like the kanban-style view of tasks. It would seem to be the perfect place to start using an AI agent - give the agent its own column / task list in Trello, anything we drop in there it deals with.

[TwistedHelixis]

Just watched this about Gems - 

 

[jmak]

2 hours ago, dhicks said:

 

Interesting - which helpdesk are you using? We are currently using Trello as a (reasonably basic) helpdesk, which currently seems to fit the way we work quite well - we have it hooked up to emails to people can email tickets in and we can reply, and I like the kanban-style view of tasks. It would seem to be the perfect place to start using an AI agent - give the agent its own column / task list in Trello, anything we drop in there it deals with.

We're using Jira Service Management - the AI agent is the built in Rovo.