What Is an Agent Harness?

You can use ChatGPT through Chat, Work, Codex and Dots. How are they different?

All four run on the same AI brain, OpenAI's GPT model. What makes them different products is the software that controls that brain, called an agent harness.

So what exactly is an agent harness?

It is the software wrapped around the AI brain. It holds hidden instructions the AI follows every time, the tools it can use, what it remembers, and the steps it repeats until a task is done. It also sets the rules, like asking for your approval before sending an email, and as a user you cannot see or edit any of it.

Then how is Chat different from Work?

Chat has a light harness, so it mostly talks with you and searches the web, which is why its answers are usually less detailed than Work's. Work gets tools for your files, email and calendar, so when you ask it to set up a workshop, it can find a free slot, book the room and invite the team.

And what makes Codex different?

Its harness gives it a workspace, like its own computer, where it writes code, runs it, spots errors and fixes them. These tasks can take many steps, so the harness keeps the key details in view and shortens older steps into a summary.

And what about Dots?

Dots are always on, with much better memory of your goals and their own computer in the cloud, connected to your apps. You can hand a dot a much longer project, close ChatGPT, and it keeps working in the background.

What to remember

GPT is the brain that does the thinking, and the harness gives it instructions, tools, memory and rules so it can finish real tasks for you. Chat, Work, Codex and Dots are four different harnesses around that one brain, which is why they work so differently.