Everyone Is Becoming a Manager of Agents and Very Few Are Ready for It

Last week, ChatGPT rolled out a significant update that separates conversations from agentic work. In the app, you’ll now see Chat and Work. They also renamed the Codex app to ChatGPT and split it into Work and Codex.

In typical OpenAI fashion, this is all incredibly confusing, especially when paired with their new model family, 5.6, which includes Sol, Terra, and Luna.

Here’s a quick overview of the changes, followed by some thoughts on what matters most and the challenges this will introduce for individuals and organizations adopting AI.

The change in structure is to accommodate new ways of working with AI. We started with simple chatbots. But as people found innovative ways to use the models, coding agents like Claude Code and Codex emerged. The labs quickly realized that coding agents are actually general agents—meaning they can theoretically do anything a human can do on a computer, as models improve and you give them the right tools.

That led to Claude Code and Codex becoming extremely powerful. Many developers and power users moved over last year. But most people don’t use them. They have been stuck with chatbots. Helpful, but with much lower impact and return on investment.

Anthropic attempted to address this by launching Claude Cowork, a more approachable Claude Code experience in their desktop app. ChatGPT Work is OpenAI’s answer to the same problem.

They are trying to make general agents accessible to a broader audience.

I think about it like this:

• Chat: Your legacy ChatGPT experience. Almost everything stays the same.
• Work: A new web interface that works similarly to Codex, their app for developers & power users. A bridge to agentic work.
• ChatGPT App (prev. Codex): The workplace app of the future, housing both ChatGPT Work and Codex.

Work and Codex run on usage-based billing, which is easy to overlook. They don’t get the same near-unlimited usage many ChatGPT Enterprise customers have grown accustomed to. This means people have to manage these tools, somewhat like we manage people, to get the most out of them. Managing agents means assigning the right work, giving them context and tools, reviewing what they’re doing, evaluating outcomes, and knowing when to step in and take over.

But few people have been trained to manage agents, and that will create problems.

To add to the confusion, OpenAI launched an excellent new model family called 5.6. There are three models in the family: Sol, Terra, and Luna.

Here’s how I think about them:

• Sol: The most powerful model. Great for hard tasks and as an orchestrator that delegates work to other agents and models.
• Terra: Good for lightweight day-to-day tasks. An affordable model for work that is clearly defined.
• Luna: A small, fast model that can be used for non-critical needs, like finding information or checking work. I mostly allow Sol to use it as needed and rarely use it myself.

That’s easy, right?

Well, each of these models has four effort settings: light, medium, high, and extra high. Sol can do most work on medium and complicated work on high. But Terra isn’t going to be very useful on anything but high or extra high. Then there’s ultra effort, which is actually a subagent workflow tool, not a reasoning or effort setting.

Completely lost? Most people are.

And I haven’t even gotten into when to use Fable, Grok 4.5, Gemini (rarely right now), or Copilot (actually some use cases now, surprisingly).

The AI labs are changing the role of employees, while leaving it to organizations to figure out how to manage the change.

And this is the problem we’re facing. Understanding how to effectively use AI in agentic workflows is a near-full-time job. You have to understand the harnesses, which general agent is best, which model will do the job well, which effort setting is right for the task, what data the agent needs access to in order to work, and how to bring along other humans for review.

This is a new way of working. And if you’re coming from ChatGPT to this world through Work, there’s so much the labs aren’t teaching. It’s an overwhelming transition.

Every organization should be investing in building workflow automations beyond individual AI. However, training remains important. Organizational transformation will take years. In the meantime, these tools are available and powerful, but people need to understand how to use them.