OpenAI launches Agents API in public beta

Since ChatGPT appeared, we have become quite good at "asking questions."

"Summarize this text."

"Tell me the issues in this document."

"Compare these three competitors."

"Write a nice email."

Until recently, these were things we would ask people to do.

Now, if you ask ChatGPT, you get an answer in seconds.

It's convenient.

Very convenient.

But lately, I've been thinking.

How long will we keep repeating this Q&A process?

Ask it to research competitors.

Look at the results and ask, "Now, put this into a comparison table."

Look at the table and ask, "Extract the three most important points."

Then, ask, "Turn that into a PowerPoint outline."

Even though it's become more convenient, humans are still sitting right there giving instructions.

It's like having a talented subordinate join the team, but you spend all day saying:

"Open Excel."

"Look at column A."

"Now column B."

"Copy that."

That's not a subordinate; that's a remote control.

That way of using AI might be about to change.

OpenAI releases "Agents API" as a Public Beta

Just looking at the name, you might think, "Another API?"

Responses API.

Agents SDK.

Agents API.

I'm already full just from the names.

But this Agents API is quite interesting if you view it not just as an API addition, but as a shift from "asking AI questions" to "handing work to AI."

A very rough explanation of the Agents API

For example, you ask ChatGPT, "Research Company A."

The AI researches it and gives you an answer.

This is the AI that many people have experienced so far.

But actual work doesn't end in one go. If the task is "Research Company A," it actually looks like this:

Look at Company A's website.

Find financial reports.

Look up news.

Research competitors.

Compare numbers.

Research additional points you don't understand.

Finally, compile it into a document.

In other words,

Work is not a question, but a "series of small tasks."

This is important.

The Agents API is closer to a mechanism that allows the AI itself to proceed with this series of small tasks.

OpenAI explains that the Agents API manages the "harness" on their side to run model calls, tool usage, context management, and sub-agents.

Moreover, this provides the mechanism that powers Codex to developers.

To put it simply, until now it was,

• Human → AI → Answer

but from now on, it will be,

• Human → Agent → Research → Judgment → Tool Usage → Re-research → Task → Deliverable

This means humans won't have to say "do this next" every single time.

This is the big difference.

Responses API, Agents SDK, Agents API. The problem of names being too similar

Let's organize this for a moment.

The names are way too similar.

OpenAI, please be a little kinder.

To simplify it immensely, here is the image:

Mechanism Roughly speaking

Responses API Call the AI

Agents SDK Build an AI agent

Agents API Run an AI agent in the cloud

Strictly speaking, there is more to it, but I think this is enough for starters.

When using the Responses API directly, developers can control the Agent Loop quite a bit, such as when to call tools, how to maintain state, and what the next process should be.

With the Agents SDK, it becomes easier to assemble agent movements using mechanisms like Agents, Tools, Handoff, Guardrail, and Session.

OpenAI itself organizes the cases for using the Responses API directly versus using the Agents SDK like this:

If it's a short process and you want to control tool execution and state management yourself, use the Responses API.

If you want to leave multiple steps, tool usage, Handoff, and Session to the runtime, use the Agents SDK. (OpenAI GitHub Pages)

And with the Agents API, we go even further.

OpenAI manages the infrastructure itself to run the agents.

In other words, it feels like you're not just renting an AI model,

but also renting a workspace for your AI employee.

AI now has a "workspace"

Personally, this is the most interesting part of the Agents API.

An agent cannot work if it only has a brain to think.

Humans are the same.

You need a PC.

You need files.

You need tools.

You also need a place to store data while you're working.

Sometimes you want to run programs.

The Agents API provides an environment where agents can handle files, execute code, and continue working while saving intermediate results.

OpenAI explains that to run agents for a long time, you need an environment that can handle files, execute code, and save intermediate results. (OpenAI)

The Agents SDK also has a mechanism called Sandbox Agent, which allows for file searching, editing, command execution, and deliverable creation in an isolated workspace. (OpenAI GitHub Pages)

This is subtle but huge.

Until now, AI was talented but didn't have a desk.

It would answer if you asked.

But once the question was over, the work was over.

From now on,

"Use this desk, use these files, and keep working until it's done"

is what we are gradually approaching.

And you don't have to do it alone

Another interesting thing is the Sub Agent.

AI can delegate work to another AI.

For example,

suppose you hand over the task of "Analyze Company A's AI strategy."

The main Agent will divide the work to other Agents:

"You handle the financial reports."

"You handle the competitor analysis."

"You handle the recent news."

It collects the results of each and summarizes them at the end.

It's quite similar to a human project team.

The Agents SDK also has methods where a central Agent calls specialized Agents as Tools, or mechanisms to Handoff to specialized Agents. (OpenAI GitHub Pages)

To put it in consultant terms,

there is a Manager Agent,

a Research Agent,

a Finance Agent,

an Industry Analysis Agent,

and a Document Creation Agent.

And finally, the Manager Agent summarizes it.

…I suddenly have a bad feeling about this.

It's a world where "AI needs a Manager, too."

Organizations will chase you even if you try to run away.

It's easy to imagine for IT system administration tasks

For example, IT onboarding for new employees.

Right now, you would ask the AI,

"Tell me the tasks required for onboarding."

Then it would tell you:

・PC preparation

・Microsoft 365 account creation

・Registration for each SaaS

・Permission settings

・Initial password notification

I see.

And then a human does it all.

So close.

What I really want is not for it to tell me, but for it to do it.

The direction of the Agents API is this.

Look at the HR system.

Confirm the new hires.

Confirm their job titles.

Determine the necessary SaaS.

Call the account creation tool.

Register a ticket in ServiceNow.

Check the PC delivery status.

Confirm if there is missing information.

Finally,

"9 out of 10 completed. 1 is on hold because the department is undecided."

If it does this, it's not an "AI that answers questions about onboarding."

It's an AI that carries out onboarding.

I think this difference is quite significant.

"It's not just about leaving everything to AI"

This is important.

Just because the Agents API is out doesn't mean we should "hand all company work to AI."

On the contrary, the more it becomes an Agent, the more things we have to think about.

What is okay to execute?

How much authority should be granted?

Where should human approval be included?

Who is responsible when it makes a mistake?

What should be recorded?

What should the cost limit be?

For example,

it might be okay to leave "checking employee accounts" to AI.

But the moment it becomes "deleting employee accounts,"

the story changes.

It's not a question of whether the AI is smart or not.

It's a question of authority and responsibility.

This doesn't change even if generative AI becomes an Agent.

In fact, it becomes even more important.

It's not just "questioning skills" that will become important from now on

When generative AI started to spread, people often said, "From now on, prompt skills are important."

Of course, how you ask questions is still important.

But as Agents increase, I feel that slightly different abilities will be required.

That is,

How to hand over work.

What the goal is.

What is okay to do.

How much they can judge for themselves.

Who to ask when in trouble.

What constitutes completion.

This is almost the same as delegating work to humans.

In fact, what good managers have thought about when delegating work to subordinates will be necessary for AI as well.

Therefore, in the Agent era,

rather than "prompt engineering becoming difficult,"

it might be that management becomes necessary.

Is the era of asking ChatGPT questions really coming to an end?

Of course it isn't ending.

"What is this?"

"Summarize this."

"Fix this text."

These ways of using it will continue to be common.

Just as mental arithmetic didn't disappear when the calculator was introduced, AI that answers questions will remain.

However, the focus of AI is shifting little by little.

Until now, we received answers from AI.

From now on, we will have AI complete tasks for us.

This difference is significant.

And I think the Agents API is a service that has shown us that change quite clearly.

From "what to ask" to "what to delegate"

It has been a few years since ChatGPT appeared.

We have always been thinking,

"What should I ask the AI?"

Writing prompts.

Crafting questions.

Looking at the answers that come back.

Asking again.

But perhaps the nature of the inquiry itself is about to change.

Not what to ask the AI.

But what tasks to delegate to the AI.

And more importantly, which tasks humans will continue to hold onto.

Looking at the Agents API, while it is technically a new API, that is not what is truly interesting about it.

It is not about "what AI can do,"

but I believe it is a technology that redraws the boundary between human and AI work.

The era of asking ChatGPT questions probably won't end.

But,

the era of sitting next to ChatGPT all day,

constantly giving instructions like,

"Do this next,"

might be starting to come to an end.

*The Agents API is a feature released as a Public Beta as of September 10, 2026. Specifications and provided features may change in the future.