OpenAI’s Dots and the Shift From AI Assistants to Persistent Agents

There is a practical difference between asking an AI to do something and giving it responsibility for an area of work. Imagine a company trying to understand customer feedback. A chatbot can summarize the latest comments. An agent can go further by grouping complaints, comparing them with previous reports and perhaps creating a list of issues for the product team. A persistent agent could instead be told to keep monitoring customer feedback, identify patterns as they emerge, investigate recurring problems, prepare small fixes when possible and escalate larger decisions when they require human approval.

The first two examples are jobs that eventually finish. The third one does not have a natural end point. It is closer to a function inside an organization.

That is the idea behind Dots that matters most. OpenAI is not simply trying to make an agent capable of doing more things during a session. It is trying to make the session itself less important. If the agent remembers the project, understands the standing objective, retains access to the relevant tools and can continue acting over time, then the user no longer needs to reconstruct the work every time they open the interface. The AI becomes part of the process rather than a tool temporarily brought into it.

This matters because real work is rarely organized into clean, isolated requests. A product launch may last several months and involve changing deadlines, competitor announcements, media reactions, sales results and internal discussions. A PR team does not simply “analyze media coverage” once; it watches coverage continuously, notices when a narrative changes and adjusts its response. A software team does not fix one bug and finish; new reports arrive, code changes and priorities shift. Sales, procurement, recruiting and research all have the same characteristic: the work persists, and much of its difficulty comes from maintaining context over time.

AI systems have traditionally been poor at this kind of continuity. Even when they can perform individual tasks well, people often have to remind them what happened before, restate the goal or provide the latest material. Persistent agents are an attempt to remove some of that overhead. Their value is not simply that they can work while the user is away. It is that they can remain attached to a problem while the problem evolves.

This may ultimately be more important than another increase in benchmark performance. An AI that can perform a difficult task in twenty minutes is a useful productivity tool. An AI that can remain responsible for a process for twenty days starts to occupy a different place in the organization.

The shift from tasks to ongoing responsibility changes the nature of the instruction itself. Early generative AI rewarded very detailed prompts because almost everything the system needed had to be supplied at the moment of interaction. The user explained the background, described the desired output, added constraints and often specified each step. Better models made that process easier, and agents began to decide some intermediate steps on their own, but the prompt was still the center of the experience.

Persistent agents weaken the importance of the prompt because they accumulate context elsewhere. If an agent already knows the project, has access to the relevant files and applications, remembers previous decisions and understands what it is allowed to do, the user does not need to describe the workflow from the beginning each time. What matters more is the standing objective.

This sounds like a subtle interface change, but it points to a broader evolution in computing. Traditional software asks the user to operate the application. Generative AI asks the user to describe what they want. A persistent agent asks the user to define what it should be responsible for.

The distinction becomes clearer with a simple example. Today, someone monitoring a market might tell an AI to search a list of websites, compare recent announcements, identify significant changes and prepare a summary. In a persistent-agent model, that instruction eventually becomes much shorter: keep watching this market and tell me when something important changes. The complexity has not disappeared. It has moved from the prompt into the agent’s accumulated context, tools, permissions and memory.

That also means the design of future AI systems may be less about creating better chat interfaces and more about creating better ways to define responsibilities. Users will need to specify what outcome matters, which sources the agent can access, which actions it may take independently and which decisions should return to a human. The interface becomes less like a conversation box and more like a management layer.

This is where the comparison with human work becomes useful, although it should not be taken too literally. People do not usually manage colleagues by giving them a complete procedural prompt every morning. They assign a responsibility, provide access to information and systems, establish boundaries and expect the person to continue until circumstances change. Dots are an early attempt to give AI something closer to that operating model.

The more responsibility an agent receives, the more important its boundaries become. A chatbot that writes a poor summary creates a small problem. A persistent agent that can edit files, interact with authenticated websites, send messages, modify code or make changes inside business software can create a much larger one. For that reason, the most consequential part of persistent agents may eventually be less visible than the agent itself: the permission and control system around it.

This is already visible in the way OpenAI describes Dots. Their actions are constrained by user and organizational permissions, and companies can determine which tools, applications and forms of computer access are available. Some actions can be configured to require human approval. These details sound administrative, but they address a central problem in agent design. As models become more capable, the question is no longer simply whether an AI can perform an action. The important question is whether it should be allowed to perform that action without asking.

Organizations already solve this problem for people through roles and authority. An employee might be allowed to prepare a contract but not sign it, approve an expense up to a certain amount but not beyond it, or change code without being able to deploy it directly to production. AI agents will need similarly granular boundaries. A communications agent might be allowed to monitor press coverage, prepare briefing notes and draft responses, but not send an external statement. A software agent could investigate bugs and create pull requests while leaving the final merge to an engineer. A procurement agent might compare vendors and negotiate terms while requiring human approval before a purchase is made.

The need for such limits becomes even clearer when the agent operates for long periods. A normal chatbot failure is usually visible immediately because the user is present. A persistent agent may make a small incorrect assumption and continue building on it for hours. The risk is not necessarily one dramatic failure. It is the accumulation of small decisions made under the wrong interpretation of the goal.

That makes reliability more complicated than accuracy. A long-running agent needs to know when it is uncertain, when new information contradicts an earlier assumption, when an instruction has become outdated and when it should stop and ask for help. It also needs to leave behind enough information for a human to understand what it has done. If supervising the agent requires reading hundreds of low-level action logs, the system has failed in another way: it has transferred execution to the AI while leaving all of the management burden with the user.

The difficult problem, then, is not simply making an agent that can succeed once. It is building one that can continue succeeding when the environment changes, exceptions appear and the original instructions no longer perfectly describe the situation. That is why persistent agents are likely to depend as much on permission design, monitoring, escalation and recovery as they do on model intelligence.

Dots are also revealing because they show how the competitive center of AI is beginning to move. For several years, the industry has been obsessed with foundation models. New releases were compared through reasoning benchmarks, coding tests, context windows, multimodal abilities and inference prices. The strongest model was often treated as the strongest product.

Persistent agents require a much larger system around the model. They need memory so that work can continue across time. They need identity so that users know which agent is responsible for what. They need access to applications, browsers or computer environments. They need permission systems, scheduling, monitoring and ways to communicate with users when something changes. They also need a mechanism for recovering when a website fails, a credential expires or a workflow reaches an unexpected state.

Once all of those pieces matter, the foundation model becomes only one part of the product.

This has important strategic consequences. If users begin assigning ongoing responsibilities to agents, the company that operates those agents can become a layer between the user and a large number of existing software products. Instead of opening ten applications to complete a workflow, the user may increasingly interact with one agent that operates those applications on their behalf.

Consider travel. A complicated trip currently involves airline websites, hotel platforms, email, calendars, maps, expense systems and messaging apps. Most of the work consists not of making one decision but of keeping all of these systems synchronized as things change. A persistent agent could eventually watch the itinerary, notice a flight change, understand which hotel or meeting is affected, prepare alternatives and ask the traveler only when a meaningful decision is necessary.

The same pattern can apply to business software. A manager may not need to open a dashboard every morning if an agent is already watching the relevant metrics and understands what counts as unusual. A communications team may not need to manually search for every new article if the system knows which companies, products and narratives matter. An engineer may not need to inspect every bug report if an agent can classify routine issues and surface the ones that require judgment.

Applications would not disappear, but their role could change. For decades, software companies designed interfaces primarily for humans. Menus, dashboards, forms and navigation systems exist because people need a structured way to operate digital systems. Agents do not necessarily require the same interfaces. They can increasingly interact through APIs, browsers and computer-use systems. If the agent becomes the primary operator, the application can become infrastructure sitting underneath it.

That is a much larger ambition than building a better chatbot. It suggests that the long-term contest in AI may be partly about who controls the layer where work is delegated. The value of that layer comes not only from answering questions, but from understanding ongoing goals and coordinating many pieces of software around them.

This leads to the economic question behind Dots.

The first wave of generative AI has mostly been understood in terms of productivity. A programmer can produce code faster. A marketer can create a first draft faster. An analyst can process more information. In each case, AI accelerates work that a person is still actively performing.

Persistent agents introduce a different mechanism. If the system can reliably own a recurring process, the gain is no longer only that the human performs the same work more quickly. Some of the work no longer requires continuous human attention at all.

That is a much more significant organizational change.

A company experimenting with ChatGPT might ask how its customer support team can become more productive. A company deploying persistent agents may eventually ask which parts of customer support actually require a person to remain involved. The first question improves an existing workflow. The second can redesign it.

This does not mean autonomous agents are about to replace entire departments, and Dots themselves are still an early product. There is a considerable difference between a convincing demonstration and a system that can be trusted with an important business process every day. Real working environments contain ambiguity, outdated information, exceptions and conflicting priorities. An agent that performs a workflow correctly one hundred times may still fail on the hundred-and-first when it encounters something it has never seen before.

But Dots make the direction easier to see. The important shift in AI is no longer only from weaker models to stronger ones, or from chatbots to agents capable of taking actions. It is from software that waits for the user to software that remains attached to a goal.

If that model becomes dependable, AI will stop being something people repeatedly call into a workflow. It will increasingly become part of the workflow itself.