OpenAI Dots: delegation is a product

Swapping two photos and publishing a website sounds like a small assignment. In OpenAI's Dots presentation on 29 September 2026, that sequence captured the pitch: agents that keep working with your context and check in when needed. For product teams, the interesting part is the agreement underneath: what you delegate, what you review, and which decisions remain yours.

OpenAI introduced Dots and ChatGPT Space at DevDay alongside models and developer tools. Dots is the personal agent; Space is the shared place where people and agents can work together. Together, they propose a shift from opening a conversation to solve a problem to leaving an ongoing responsibility in someone's hands.

Our read is that this transition needs as much design as intelligence. Asking AI to prepare a change is a fairly concrete request. Allowing it to judge when that change is ready creates a different relationship with the product.

What are OpenAI Dots?

In the keynote, OpenAI described Dots as persistent agents: they can follow an assignment over time, use connected applications and work on a cloud computer. The promise is that you can ask one to monitor user bug reports, investigate problems and prepare fixes for review.

The software migration example makes the scope easier to picture. Asking an app to stop depending on an old API, its connection to another service, means finding where that connection is used, changing code and checking what breaks. In the presentation, that work led to proposed code changes ready for review. The person still had a specific role in the process.

The Agents API documentation shows part of the technical foundation: sessions that preserve work, tools and environments where agents can run code. OpenAI manages session continuity and provides ways to follow progress or intervene. Those documents describe platform capabilities; they do not establish that a Dot can handle every project well.

The launch describes a product promise. Everyday use will tell us how well Dots interpret context and how much supervision they end up needing.

Doing the work and owning the assignment

During the demonstration of Blossom, a fictional music app, the agent found feedback and a new design close to launch. The presenter could ask it to turn the design into a working version. The scene combined two jobs: noticing that something had changed and acting on the change.

For a product team, the second job depends heavily on the first. A freshly shared design could be a proposal, an experiment or an approved decision. Access to the file does not settle that distinction. The assignment needs to establish which sources carry authority and how far the agent can proceed.

Imagine delegating bug monitoring for your checkout. A useful agreement might allow investigation and a proposed fix, require purchase-flow tests and leave deployment to someone on the team. If the agent discovers that the fix would change prices, the task now needs a different decision. That boundary belongs in the experience you design.

The official Dots announcement adds an important detail: proactive background research uses read-only tools. Custom rules let users permit actions, require approval or block them. OpenAI also provides an activity view for following the work. Being available around the clock does not mean having permission to change everything.

Approval needs to support a decision

An approve button does little if you first have to reconstruct what the agent did. To publish those two photos, you would want to see which images changed and how the page looks. For a code fix, you would want to know which behaviour was tested and what remains unchecked.

At Labba, that is where we would put the design effort: making the result arrive with enough evidence to support the next decision. The initial permission and final review need to connect. If the request was to change images, the approval should not conceal a modification to the purchase form.

OWASP recommends minimum permissions and human approval before high-impact actions. It also says authorization should be enforced by the systems executing those actions. Our product conclusion is that these controls must be understandable to the person delegating. A rule nobody knows how to configure is unlikely to make work feel dependable.

Where ChatGPT Space fits

According to the DevDay recap, Space brings together shared knowledge and pages where people and agents can collaborate. In the keynote, teammates mentioned a Dot inside a page to request changes and work on the same material.

The potential value becomes clearer when delegation involves more than one person. If you request an update and a colleague reviews the result, both of you need to understand the assignment. A private message saying the work is finished leaves too much interpretation to everyone else.

Our bet is that a shared page should help distinguish proposals, approved decisions and unfinished work. Space provides a place for that collaboration; practical use will show how well it handles those distinctions. An agent-maintained document can be current without representing a team agreement.

The NIST framework calls for defined responsibilities and oversight in AI systems. Our reading is that this clarity should also appear where the team works, visible to the people assigning and reviewing tasks.

What to try first

OpenAI is rolling Dots out in eligible markets for Pro and Business Premium; Enterprise beta access requires an administrator to enable it. The announcement also distinguishes conversations with a Dot from tasks it starts in Codex or ChatGPT Work: those tasks still count toward the relevant usage limits.

We would start with a bounded responsibility whose output you can review, such as investigating bug reports and preparing proposals. The reason to expand that assignment would be how much follow-up work it removes without leaving you guessing what happened.

The revealing test will come when a priority changes halfway through an assignment. If a clear instruction is enough to redirect the work, and the team understands what remains unfinished, Dots will have earned something more useful than another demo: a stable place in how work gets done.

Sources

• OpenAI DevDay keynote, 29 September 2026: transcript supplied by Manuel; website publishing, migration and Blossom examples.
• Introducing dots, OpenAI, 29 September 2026.
• DevDay 2026 Recap, OpenAI, 29 September 2026.
• Agents API, OpenAI documentation, accessed 29 September 2026.
• Excessive Agency, OWASP, 2025 edition, accessed 29 September 2026.
• AI RMF Core, NIST, 2023, accessed 29 September 2026.