If you set up a Custom GPT a year ago and haven’t really used it since, I suspect you already know something is off. It felt exciting when you built it. It sat there. You navigated to it occasionally. You told yourself you should use it more.

You were not wrong about the idea. You were just ahead of the implementation.

Custom GPTs were a promise. Skills and workspace agents are, finally, the delivery.

What Custom GPTs were actually good at

It is worth being fair about this. Custom GPTs were genuinely useful for a period. They let you shape how the AI behaved, add reference documents, and create something with a consistent persona for a specific task. A lot of people built them for customer-facing use cases, for internal reference tools, for writing in a particular style.

The problem was the friction. You had to navigate to the GPT specifically. You had to remember to use it. You could not combine it with anything else. If you wanted the GPT’s knowledge and GPT’s general capabilities in the same conversation, you were doing a lot of manual switching. And updating a Custom GPT when the underlying documents changed was, charitably, clunky.

They were useful. They were also always slightly more effort than they should have been.

What skills change

Skills in ChatGPT work differently in a way that sounds small until you experience it.

A skill is a modular unit of expertise. You build one skill for one thing: writing in your voice, generating images in a specific art style, formatting a particular kind of document, producing LinkedIn posts in a set structure. Each skill has a description, a set of instructions, and reference files. The AI knows when it is relevant and loads it automatically, or you call it explicitly with an @ mention.

The critical word is stack.

With Custom GPTs you had to pick one. With skills you can combine them. If I am writing a blog post, I can call my voice skill, my image style skill, and my social media post skill into the same conversation, and they work together without getting in each other’s way. Each skill does one thing well. The combination does something that would have taken twenty minutes of manual tab-switching a year ago.

Testing this properly is worth doing. Take a story from a BBC news page - something you have no professional interest in, horse racing, or the results of a county cricket match - paste it in, and ask the AI to rewrite it in your voice using your skill. Read it back and note where it jars. That is where your skill needs refining. It is a much better calibration exercise than testing with content you would naturally talk about, because you can actually spot the seams.

Once a skill is working, you can download it as a zip file and install it in Claude too. The same skill, working across both tools. That cross-platform portability is something Custom GPTs never offered.

What workspace agents change

If skills are the expertise, workspace agents are the workflow.

A workspace agent is a configured automation that can run on a schedule, access real tools - email, calendar, Slack, SharePoint, CRM - and carry out a sequence of tasks without being asked at each step.

I have a morning briefing agent that reads my inbox and calendar every weekday at 6.30am and sends me a prioritised brief. It does not ask if it is okay. It reads, reasons, and sends. I gave it permission to send emails and that is the only destructive action it is allowed to take. Everything else is read-only. That constraint is deliberate and important. You are setting the guardrails. The agent operates within them.

The more interesting version is the content agent I built that chains three skills together. When I have finished a piece of research or a draft, I tell it to do everything. It takes the raw content, applies my voice skill, generates three image options in my art style, creates a set of social media posts, and packages the whole thing. Five minutes later, it is done. I have not switched a single tab.

That is not a chatbot assistant. That is a production workflow.

Why Custom GPTs cannot do this

Custom GPTs can do some of this in a limited way. But they cannot be scheduled. They cannot act autonomously. They cannot combine the expertise of multiple skills. They cannot connect to live tools in the same integrated way. And they are attached to a persona rather than a process.

OpenAI itself has positioned workspace agents as the successor to Custom GPTs, not an addition to them. That framing is deliberate and accurate. If you are still thinking in Custom GPT terms, the mental model needs updating.

Claude’s equivalent is different but worth knowing. Claude does not have workspace agents in the same scripted, scheduled sense. What Anthropic calls “managed agents” are agents that Claude spawns autonomously when it is handling a complex task, delegating sub-tasks and orchestrating them rather than following a user-defined workflow. Claude’s Cowork mode provides something adjacent for local machine access: give it a folder, and it can read, edit, and create files in that folder continuously, which is powerful for project-based work. But for the scripted, scheduled, autonomous workflow? ChatGPT’s workspace agents currently have the edge.

The practical question

The useful question is not “what is the difference between all these things?” It is “what repeatable job in my week could an agent handle?”

Think about: the briefings you prepare before client calls, the weekly reports you pull together from multiple sources, the social media posts that follow every piece of content you publish, the inbox triage that eats the first thirty minutes of your morning. Those are the starting points.

Pick one. Build the skills that support it. Define the steps clearly. Set the permissions explicitly. Test it on something low-stakes. Let it run.

The Custom GPT era was about shaping the AI’s personality. This era is about shaping the AI’s workflow. That is a more useful thing to do.

Sources: Introducing workspace agents in ChatGPT (OpenAI, 22 April 2026), OpenAI Launches Workspace Agents (April 2026)

Scott Quilter | Co-Founder & Chief AI & Innovation Officer, Techosaurus LTD