Episode 140: Why ChatGPT Workspace Agents Are Bigger Than Custom GPTs

It's just like there's so much you can do and the power actually comes from the integrations, the memory, and the skills. It's amazing, like you have someone that actually does the work for you. You just share some connections with them, like your products you use, then tell them instructions and they do it for you. But this time it's an AI Agent.

Hi, my name is Demetri Panici and I'm a content creator, agency owner, and AI enthusiast. You're listening to the AI Agents podcast brought to you by Jotform and featuring our very own CEO and founder Aytekin Tank. This is the show where artificial intelligence meets innovation, productivity, and the tools shaping the future of work. Enjoy the show.

Hello and welcome back to another episode of the AI Agents podcast. In this episode, we're going to be talking about ChatGPT workspace agents, a new thing that GPT has recently released. These are cloud-hosted agents inside your ChatGPT and are a big upgrade from what custom GPTs were before.

How you doing today, Adigan?

Hey, Demetri, how are you doing?

Doing good. Yeah, I've been busy. We are very busy at Jotform building all these AI tools every day. We're building all these workspace features. ChatGPT came out with this workspace agents feature and I think it's really good, though it's really buggy. There are a lot of buggy things but the potential is there to do so much, especially if you are doing something every day or every hour. It's just so easy to automate that work with the ChatGPT workspace agents.

That's why we wanted to cover that today. Go ahead, start.

I'll kick things off by showcasing what they've given us as an example, a good video to kick things off and their kind of opening web page. Here it is, give me a sec.

So this is basically something that has probably been in the works for a while. It's cloud-side agents that can actually do things and interact with things and use different MCP connectors and whatnot. We were alluding to this a while ago with MCPs, how there's going to be so much difference in connectability. In the beginning, it was very active; you had to chat with stuff and interact with the MCPs, but now it's a little bit more active on its own.

It's powered by Codex, which is one of the big things that ChatGPT has recently improved on. They can take on a lot of tasks that people already do for work, like preparing reports, responding to messages, writing code, and they all run in the cloud. There are a lot of really good examples here, like software reviewer, product feedback router, weekly metrics reporter, lead outreach agent, third-party risk manager, and much more.

You'll see a lot of examples that people use. It's basically what you do for your daily tasks that you don't want to end up doing. Anyone who's ever put together a weekly report knows a lot goes into it, but practically AI can do that report very easily for you. What's interesting is they only released it to their business plan. Even though I pay for ChatGPT Pro, which is personal, I don't have access to it. So it's definitely more of a team-based thing.

It's interesting to see them release this because there have been some other products that have come out recently which I can compare and contrast a little bit. But first, how is this different than what the product was prior with custom GPTs? Custom GPTs couldn't run when you were logged off, there were no scheduled capabilities, you couldn't interact with Slack, which is a big one. There was limited code execution with Codex, no multi-step workflows with memory, no approval gates on writes, no shared workspace capabilities, and the skill standards were worse across the board. It's a big improvement from custom GPTs.

What makes this so powerful is three things: integrations, skills, and memory. First, integrations. ChatGPT added an app store recently with many great apps like Gmail, Google Drive, Slack, Jotform, and more. The agents can now use these products. Second, skills. You can create skills now, which are basically files describing workflows or tasks. If you do something every day, you describe it, and if the agent makes a mistake, you update the file. The agent can even update those files itself. Skills are just combinations of files that describe your workflow.

Third, memory. The agent can remember stuff. You can give it a list or database like a spreadsheet, and it will remember all past information and use it. The combination of these three things makes it so powerful. Everyone is talking about how SaaS is over, but I feel like this is just making SaaS so much more powerful because now you can easily have someone using your Gmail, Google Drive, Slack, and the AI is basically doing the task for you. When you give access to your calendar or Slack, it can send you messages or put things on your calendar. It's able to do the work for you.

What's great is it's running on the cloud, so you can say, 'Hey, I created this agent, it should run every hour,' and even if you shut down your computer, it will still run every hour because it's running on ChatGPT's server. You don't have to keep your computer open, which makes this very powerful.

We have great examples, but do we want to compare to Claude's managed agents?

Yeah, for sure. There's another product Claude recently released called Cloud Managed Agents, which is very similar to what they have here.

Cloud Managed Agents is a suite of APIs for building and deploying agents at scale. You define agents with specific tools, personas, and capabilities. You configure sandbox environments with the right packages and network controls. You fire off sessions from your own application, and Claude does the work inside an isolated container with full file system access, bash execution, and web search.

I have a Kanban board sitting on top of Managed Agent. I drag one over to in progress and that fires off a session automatically. The ticket says optimize website performance, so my backend creates a session pointing to an environment configured with Lighthouse and Puppeteer pre-installed and mounts my GitHub repo into that container. Claude has the code base, but it's definitely not made for normal people. It's a little more advanced, made for developers.

The ChatGPT Workspace Agents are definitely made for normal people. There's no technical jargon like this. That's the main difference; Claude is more developer-oriented, while ChatGPT agents are for normal users.

Should I go ahead with my demo?

Yeah, for sure.

Here I am on ChatGPT. It says developer mode, but I guess I enabled some stuff. The agents are not available on regular ChatGPT; you need a business, education, or enterprise account. I'm showing from our enterprise Jotform account. If you think about it, you're either a business to use these automations. You can get a two-person business account, which is cheaper than the max plan, so it's not unreachable but a bit more expensive than regular or free ChatGPT.

It starts showing these agents here. I'm going to create two examples I already created because it takes a long time. I'll show how to create them as well. Let's start with the Jotform uptime monitor agent. Say you have a small business with a website and want to be aware if your website goes down. Create an uptime monitor agent that checks a list of URLs every hour and emails you if any pages are down. It notifies you if your website is down and ensures it's always up.

It's similar to ChatGPT; you just type it here. I realized I need to create it as an agent, so let me start from the agent part. Create an agent by clicking the plus sign. This is the agents page where you can pick from templates. The agents you build are listed here or in your business or organization if others share their agents. You can reuse or duplicate them. For example, the uptime monitor agent I created today can be used by other Jotform employees if shared.

I'm going to create a new agent. The create agent page is really cool, with an icon and name. I'm typing: Create an uptime monitor agent that checks a list of URLs and emails me if any pages are down. That looks good, so I submit it. It starts building the agent inside the agent builder. You don't have to use the builder directly; ChatGPT builder builds the agent for you. Once built, you can customize it manually, similar to how Jotform AI creates forms that you can edit manually.

It's checking whether the available app can support emails. It says the agent will check the list of URLs, use Gmail, run every hour, check website uptime, detect outages, and email incident alerts. I like everything and start building. The agent builder shows the chat part and the agent builder. The agent is building itself, and I can see how it's making changes. For example, it adds instructions for URL monitoring and email alerts, sets the schedule, and adds a description.

While it's working, I can also edit it myself. Both the AI and I can work on it simultaneously. It's almost finished but needs the URL list to check if it's up. I'm giving it the URL now.

I mentioned three important things the agents have: integrations with apps like Gmail, Slack, Jotform; skills which are files describing workflows that are reusable and transferable across platforms; and memory that remembers past information. The schedule is set to run every hour to check the URL uptime. I can manually edit whatever it is doing. They built an agent builder and an agent that builds agents, which is pretty cool. The agent is ready to try in chat.

It's pretty fast, so I don't have to show the one I created previously. I check if the website is up. The agent activates the computer. Since these run on the cloud, they need to activate something every hour. Let me show the one I created previously. It runs multiple times, some manually and some on schedule. I can see analytics and previous runs, though I'm still learning how to find them.

The agent checked jotform.com; it is up and loaded the homepage. I can see the activity and previous runs on the agent page. The second agent I'm going to create checks Google Docs files every day and reports any changes by email to my demo email address. It needs Gmail and Google Docs integrations. Usually, providing the URL is enough.

While this is working, I'll show the one I created before. It has Google Drive and Gmail integrations and memory with the Google Docs file. This is the first version of the file. I'll add something to create a second version and see if it catches the new addition. I'll run it now and see if it works.

The agent boots up the computer. This uses credits; I think it was free until yesterday, but now it probably charges credits. You want to be mindful of running it too many times as it might cost a lot. The agent is checking the document, loading the monitor, reading the current Docs file snapshot, and sending the baseline change email if needed.

Because there was no previous version, it didn't detect changes but is sending an email now. It asked for approval, probably because it was the first time. Monitoring started for the test Google Docs file. It saved the baseline locally in memory. The memory section on the app shows the saved content.

Let's check if I received any emails. Monitoring started for my test Google Docs file with no prior saved versions. I'm going to change the file to the third version and run it again to see if it discovers and reports changes.

The agent is showing here. I can disable it by clicking on it. This is a new UI that shows if you are talking to an agent, not just planning. It monitors Google Docs compared to the previous version and always asks for approval. You want to run the agent a few times yourself, perfect it, and then run it hourly or daily as preferred. It detected changes and will send me an email.

Let's check if it sent the email. Refreshing the inbox shows an email titled 'Changes detected: Edit third version.' It discovered what changes were made. This is very useful if you share a Google Docs file and want to be triggered whenever someone makes a change. Google Drive already has this feature, but this is just an example. You can receive an email or send it to Slack. I was working on the Slack integration but had some problems, so I can't show that demo. This is amazing; you can do so much with this.

It's interesting because many new platforms went from requiring a lot of work into prompting and making agents that had to be manually worked with, to now where you can just tell it what you want to do and it does the research and executes it for you. We're working in a totally different landscape than six months ago. It's amazing.

The power comes from the integrations, memory, and skills. It's amazing, like having someone who actually does the work for you. You just share some connections with them, like your products you use, and give instructions, and they do it for you. This time it's an AI agent. You could do this six months or a year ago but only with cloud code or cursor, which required writing code. Today, any normal person can get a business or education account and start building this stuff. Amazing.

I'm curious, for what you're doing with Jotform, what is the long-term vision for different AI agents and how you're trying to implement stuff like this? How does it inspire you to improve your own AI suite?

These days, we are working on Jotform's own workflow features which are very similar. The question is, this one already has so much, including Jotform integration, and can do many things. The main difference is nailing down things. You can build all the stuff we are building with Jotform AI, but you have to spend a lot of time trying to get things right. We provide already optimized and opinionated features and products, as opposed to building them from scratch.

Also, you have to learn this stuff and have a ChatGPT business account, which with Jotform you don't need. Even with a free Jotform account, you have these features with some limits. You don't need API keys or ChatGPT connect. All these AI features come with Jotform already optimized and correctly executed. That's the main difference. These are amazing and serve as inspiration for us to do similar things within our product, which is exciting.

There is a worry that SaaS products like Jotform or others can be built using Codex or cloud code or agents that do the same things. If you have a simple product, maybe you can, but established products have been building for a long time and made mistakes. Building your own SaaS means making the same mistakes and it might cost more to host and run yourself.

There is a balance. If you want stuff sold by fast products, it's cheaper, easier, stable, less buggy, and more secure to let managed SaaS products handle it. But if you're integrating all these SaaS products or building custom workflows, use products like ChatGPT Workspace Agents to duct tape them together and automate your work. This saves time to focus on important things instead of repetitive tasks.

Thanks for the insight and showcase. Sorry I don't have the business edition, but I appreciate you holding down the fort and showing things. Let's keep covering cool stuff. See you next time.

Thank you so much for watching and we'll see you in the next one.