I demoed ChatGPT Agents this morning in an AI Operators meeting with one of our clients. The room had about forty people on it, most of whom have spent the last few months getting comfortable with ChatGPT, building their first skills, and learning how to brief the model with the right context. A few have started experimenting with agents on their own. Most have heard the word and weren't sure what was different about it.
I want to walk through what I showed them, because the feature is more than a name change. ChatGPT Agents are doing something the previous generation of tools couldn't, and the differences matter for how teams should think about the next layer of AI work.
A quick anatomy
An agent, in this context, is a piece of software wrapped around a large language model that does four things in a loop. It reasons about what you asked, it makes a plan, it takes actions using tools, and it produces an output. The loop runs as many times as needed until it decides the work is done.
That description sounds like the chat experience most people already use. The difference is what the agent has access to during the loop. It can call tools like web search or code execution. It can hit connected applications like SharePoint, Jira, Slack, GitHub, or Canva. It can pull from documents you upload as knowledge. It can read and write its own memory. And it can run on a schedule, not just when you open a chat window.
That last point is where things start to get interesting.
Memory is the part nobody is talking about correctly
For the past year, people have been telling me their ChatGPT was learning about them. They'd say something like, "the more I use this project, the better it gets." That wasn't quite right. Custom GPTs and projects had access to the files and instructions you gave them, but the model wasn't actually retaining anything between sessions. It was reading the same context every time and making it look continuous.
Agents are the first version of this where memory is real. After an agent runs, it writes markdown files into its own memory directory based on what it learned. I showed the room the memory folder for my Chief of Staff agent. Inside it had a file called "relationship context" where the agent had started building an ontology of our clients, who works on what, and how they relate to each other. It had another file called "open loops" with a list of decisions I needed to make, pulled out of conversations and meeting transcripts.
I didn't tell it to make those files. It made them because it decided they would be useful to it the next time it ran. Every time the agent runs, it updates those memories. Over time it gets a clearer picture of the work it's helping with.
This concept does not exist in custom GPTs, projects, or skills. It is genuinely new for ChatGPT users, and it changes the trajectory of what an agent can do for you over weeks and months.
Scheduling makes them actual workflows
The other piece I want to flag is the schedule. From the same edit screen where you configure your agent, you can tell it to run on a recurring trigger. Every morning at 7am. Every Sunday night. Every two hours during business hours. You can stack as many schedules as you want.
That sounds small. It isn't. It moves the agent from a thing you go talk to into a thing that operates whether you're at the keyboard or not. A morning brief that pulls open threads from email, calendar, and Slack and lands in your inbox before you start the day. A Sunday night summary of next week's commitments. A weekly competitive scan that reads the news and writes you a memo on what changed.
This is also where agents start to overlap with the workflow automation category that has existed for a long time. Zapier, n8n, Power Automate, and Alteryx have always been able to move information from one system to another on a trigger. Then they added AI as a step inside their flows. ChatGPT and Claude are now doing the inverse, putting workflow automation inside the AI tool. The two categories are converging, and the AI-native version has the advantage of reasoning natively about what to do at each step rather than running fixed branches.
For most teams, this is the moment where individual productivity gains start to look like real workflow automation.
The composition is the third thing
When you open the edit screen for an agent, you're configuring four things side by side. You're choosing which connected apps it can reach. You're attaching the skills you want it to use. You're uploading any knowledge files it should reference. And you're writing the prompt that tells it what its job is.
That arrangement matters. In the past, getting an AI tool to do useful work meant either configuring a custom GPT, or building a multi-step Zapier flow, or writing a prompt long enough to capture all the context. Agents collapse those options into one place. The skills hold the writing style and working rules, the connected apps bring in the live data, the uploaded files act as reference material, and the instructions carry the assignment itself.
I've used the phrase "conductor of knowledge and context" in the AI Operators meetings for a few months now, mostly to describe what good AI users actually do. Agents are the first feature inside ChatGPT where the conductor metaphor maps cleanly to the interface. You can see the parts you're orchestrating. You can adjust them. You can hand the agent more context when it needs it.
A note on credit
Claude got to this territory first. Anthropic's scheduled tasks in Cowork and routines in Claude Code have been running on this model for a while, and some of the patterns are more mature there than they are in ChatGPT. You can also build those Claude tools conversationally, so the conversational-creation piece itself isn't what differentiates ChatGPT. What ChatGPT did well is wrap the same primitives in a more polished and manageable experience. The agents you create show up in a clear list, the edit screen surfaces the moving parts in plain language, and the templates on the create page guide first-time users into reasonable starting points. Claude has historically optimized for the power user, which gives Anthropic an edge with technical teams but leaves a real gap for everyone else. ChatGPT productized that gap.
If you're inside a large enterprise that has standardized on ChatGPT, this is the version you'll be working with. If you're choosing between platforms, the underlying capability set is now close enough that the decision comes down to your existing connectors, your governance posture, and your team's comfort with each interface.
Full disclosure
The Gadoci Consulting team is still heavily a Claude shop. Most of our internal work, our skill libraries, and our custom builds live there, and we've been recommending Claude to clients with technical depth for over a year. What changed for us today is that ChatGPT got slightly interesting again for the first time in a long while. The agents feature, combined with skills and apps, finally adds up to a platform story we'd want to point clients toward in certain situations.
The caveat: agents only exist in the ChatGPT web experience. A handful of the more useful recent additions are web-only too. That gap between what's in the browser and what's in the desktop app is a real signal of how far behind the desktop product is, and it's worth knowing about if you've been treating the desktop app as the primary surface.
We're hopeful that the competition between Anthropic and OpenAI keeps producing tools at this pace, and we're grateful for the work both teams are putting in. The people we serve benefit either way.
What this changes for AI Operators
For anyone trying to move AI from individual use into team and department workflows, agents lower the activation cost in three real ways.
The setup is approachable for non-technical users. You can build a working agent by talking to ChatGPT for ten minutes, and once it exists you can find it, edit it, and rerun it without hunting through configuration files. You don't need to know how to wire a Zapier flow or sit inside a developer tool.
The composition is reusable. The skills you build for one agent can be reused across others. The connectors you enable for one team carry over. Once a team has a few good skills and a few good connectors, the cost of building the next agent drops significantly.
And the memory layer compounds. The agent gets better the more you use it, because it's writing down what it learns. That changes the math on whether to invest in an agent for a recurring workflow, because the second month is materially more useful than the first.
This is the closest thing I've seen to operationalizable AI inside ChatGPT itself. It is not finished. There are kinks. The connector list depends entirely on what your IT team has approved. Memory is rough around the edges and will need attention in production use. But the pieces are there, and the pieces matter.
If you're running an AI program inside an organization, this is a feature worth understanding before your operators start asking about it. They will ask. The right move is to be ready with a few good agents already built and a clear point of view on where they fit in your workflow stack.