OpenAI to Acquire Ona: Codex Is Becoming a Cloud Workspace for Long-Running Agents

# OpenAI to Acquire Ona: Codex Is Becoming a Cloud Workspace for Long-Running Agents

OpenAI has announced plans to acquire Ona. According to OpenAI, the acquisition will expand Codex with secure, customer-controlled cloud infrastructure for long-running agents across software and knowledge work.

This update matters because it shows a major shift in AI product design:

AI is no longer only generating answers or code. It is starting to work inside persistent cloud environments.

In the past, users often used AI coding tools inside a local IDE, web chat or terminal. The user asked for code, copied the result, tested it, pasted back errors and asked the AI to fix them.

But if Codex gains a more persistent cloud workspace, it can support longer and more complex tasks. It can keep context, maintain task state, continue across sessions and return reviewable results.

That is a different kind of AI product.

Codex needs a persistent place to work

A normal chatbot follows a simple pattern: user input, model output.

An AI agent needs more than that. It may need to:

• read project files
• understand a task
• modify code
• run tests
• remember intermediate state
• wait for user confirmation
• continue later
• produce a final deliverable

These tasks are difficult to complete in a single conversation.

The strategic value of Ona is that it gives Codex a more reliable execution environment. The goal is not only to make the model smarter. The goal is to give the agent a place to work.

This points to the next stage of AI workspaces. Users do not want a chat that restarts from zero. They want an AI workspace that can continue work over time.

Cloud agents change the coding workflow

Traditional AI coding workflows require constant human involvement.

A typical flow looks like this:

1. The user describes a task. 2. The AI generates code. 3. The user copies the code. 4. The user runs tests. 5. The user pastes errors back. 6. The AI tries again.

This works for small tasks, but it is not ideal for larger projects.

A cloud agent changes the workflow.

The user assigns a task, and the agent works inside a cloud environment. It can preserve files, context, execution state and intermediate results. Then it returns progress or final output for the user to review.

This makes AI coding feel more like real collaboration instead of one-time code generation.

Possible tasks include:

• fixing bugs
• refactoring modules
• adding tests
• upgrading dependencies
• writing documentation
• investigating CI failures
• preparing pull requests

All of these tasks benefit from persistence and reviewability.

Long-running agents make token cost harder to predict

Cloud agents are powerful, but they also make token cost more complex.

A long-running task is not one model call. It is a sequence of calls.

A Codex-style workflow may include:

1. reading the task 2. scanning files 3. choosing an implementation path 4. generating code 5. running tests 6. analyzing errors 7. revising the code 8. summarizing the result

Each step consumes input and output tokens.

If a task runs for hours or continues across sessions, total token usage can become difficult to estimate.

This is where a Token Calculator becomes useful. Users should not only compare model prices per million tokens. They need to estimate the cost of the full agent workflow.

The better questions are:

Is this task worth giving to a premium agent?

Can a cheaper model prepare the task first?

Which steps need a stronger model?

Should output length be limited?

Should context be compressed regularly?

Prompt Optimizer becomes a pre-flight check for agent tasks

Long-running agents are sensitive to unclear instructions.

If a user writes:

Improve this project.

The agent may not know what to improve: performance, structure, UI, copy, dependencies, tests or deployment.

Vague tasks cause agents to take unnecessary paths, generate irrelevant changes and consume more tokens.

A better agent prompt should define:

• task goal
• code scope
• areas that should not be changed
• acceptance criteria
• output format
• testing requirements
• change summary requirements
• human confirmation points

Prompt Optimizer helps turn vague requests into executable task briefs. Clearer prompts reduce retries, improve output quality and make token cost more controllable.

AI Workspace should manage execution, not just chat history

OpenAI’s Ona acquisition also shows why AI Workspace products matter.

A useful AI Workspace should manage:

• current task
• current model
• current context
• token or credit usage
• execution stage
• completed steps
• items waiting for user review
• outputs that can be saved or reused

This matters for developers, product managers, operators and researchers.

Real work is not a single answer. It is a process.

Toket AI’s AI Workspace can build around this direction: helping users manage model selection, prompts, context, cost and results in one place.

What Toket AI users should take away

OpenAI’s plan to acquire Ona shows that AI agent competition is moving from model capability to execution environment and workflow capability.

Users will not only ask:

Which model is strongest?

They will ask:

Which agent can keep working?

Which environment preserves context?

Which model fits this stage?

How many tokens will a long task use?

Is the prompt clear enough?

Which results need human review?

For Toket AI users, a practical workflow is:

1. Use Token Calculator before starting long tasks. 2. Use Prompt Optimizer to clarify the task brief. 3. Use AI Workspace to manage long-running work in stages. 4. Avoid dumping all context into the model at once. 5. Add human confirmation points for important tasks.

The next stage of AI agents is not only stronger models. It is persistent execution, clearer prompts, controlled token cost and workspace-style task management.