OpenAI dots Turns Coding Agents Into Personal Assistants, With Limits on Delegation

OpenAI dots moves work beyond individual prompts, but its promise of ongoing assistance brings a harder question: how much judgment should users delegate? OpenAI describes an assistant that continues working between conversations, remembers relevant context, and returns when it needs a decision. That changes the relationship between a person and an AI agent, software that uses tools to pursue an assigned goal.

At TechCrunch Disrupt, OpenAI executive Alexander Embiricos said, “I personally never intend to have a meeting with someone else’s agent.”

That position captures the tension behind the product. An assistant can prepare work for its owner without becoming an acceptable substitute for that owner. OpenAI and Anthropic are extending capabilities associated with coding tools into everyday work. Their products now face a broader challenge: deciding when software should proceed, when it should ask, and when another person deserves a human response.

OpenAI dots Changes Who Keeps the Work Moving

OpenAI’s personal assistant documentation describes dots as a continuing agent with its own cloud computer and browser. The company says it can research, analyze data, prepare documents, and build software. Users can reach the same assistant through supported interfaces while work continues elsewhere.

The meaningful change is the transfer of responsibility for keeping a task moving.

In a conventional chat interaction, the user usually supplies the next instruction. Even when an answer contains useful work, the person remains responsible for reconnecting it to the project. They decide what happens afterward, gather missing information, and return when another step needs attention.

An ongoing assistant changes that expectation. The user delegates an outcome and expects the software to maintain enough context to continue pursuing it. Success therefore depends on more than the quality of an isolated answer. It also depends on whether the assistant understands unfinished work.

Consider a hypothetical product launch. A chat tool might summarize feedback when someone pastes it into a conversation. An ongoing assistant has a broader assignment: follow the launch preparation, identify an unresolved dependency, and prepare material for the person responsible.

Those are different product promises, even if both systems can generate the same summary. The second promise includes continuity, prioritization, and the ability to recognize when circumstances have changed. It also creates more opportunities for misunderstanding the user’s intention.

This explains why the move from coding tools to personal assistance involves more than removing a technical interface. Software development offers relatively concrete intermediate results. A proposed change can be inspected, tested, and accepted or rejected against an existing project.

Personal work often has less explicit completion criteria. Preparing for a conversation might require understanding an unresolved disagreement rather than producing a document. Recognizing that distinction demands context about people, priorities, and the reason a task matters.

A friendly identity can make the assistant easier to address. It cannot establish whether the assistant has understood the responsibility behind a request. That distinction matters whenever the product’s presentation encourages users to treat it like a dependable colleague.

OpenAI’s opportunity is to make continued assistance understandable without requiring users to manage every intermediate operation. Its obligation is to keep the boundary between assigned responsibility and assumed authority visible. Otherwise, convenience can conceal decisions the user never intended to delegate.

Coding Tools Supply Execution, but Personal Work Adds Ambiguity

The connection to coding tools is clearest in the underlying work pattern. An agent receives a goal, gathers information, uses tools, examines results, and continues. Moving that pattern into general work expands the range of people who can delegate tasks.

OpenAI’s documentation on tasks and memory says dots can maintain ongoing responsibilities, handle recurring work, and use saved context. It distinguishes assigned tasks from proactive research, background reading intended to find useful connections. The company illustrates that research with a possible conflict between a release decision and an earlier draft.

Execution becomes more valuable when the assistant can recognize the next relevant task, but recognition is harder to verify than output.

A generated document is visible. A missed connection is often invisible until someone encounters the resulting problem. This creates a measurement challenge for personal assistants: users can inspect what the system produced, but they cannot easily inspect everything it failed to notice.

The distinction also changes the meaning of helpfulness. An assistant that produces an accurate summary has completed a bounded operation. An assistant that decides a summary deserves immediate attention has made a judgment about the user’s priorities.

Imagine an assistant reviewing materials for an upcoming customer discussion. It finds a product statement that conflicts with an older planning document. Flagging the discrepancy could prevent confusion, provided the older document still reflects an active commitment.

The same alert could waste time if the older document contains an abandoned proposal. The difficult work lies in establishing which information remains authoritative. Simply retrieving both documents does not resolve their relationship.

This is where ongoing assistance overlaps with personal knowledge management. Useful context requires more than collecting material. It requires preserving enough information about decisions, ownership, and timing to interpret that material correctly.

For developers, this suggests a broader role for familiar agent capabilities. An assistant might connect implementation work to feedback, planning, and review. The value would come from maintaining those relationships, rather than generating more code in isolation.

For knowledge workers, the attraction is similar. The assistant promises to reduce the effort of restarting a project after an interruption. That benefit depends on whether its understanding remains current as priorities shift.

The unresolved issue is how reliably dots identifies those changes. Documentation establishes intended behavior, not a measured success rate across ordinary workplaces. Evaluating the product requires separating the ability to continue working from the ability to continue working on the right thing.

Always-On Work Still Depends on Access and Context

The phrase “always-on” invites an expansive interpretation. Users might reasonably imagine an assistant that can reach any connected resource whenever necessary. The documented operating model is more specific.

OpenAI’s computer access documentation distinguishes the assistant’s cloud computer from a user’s connected computer. Cloud work can continue while the user’s machine is off. Tasks requiring that machine depend on its availability and the desktop application remaining open.

A persistent assistant still operates within the reach of its tools and connections.

That distinction changes how buyers should interpret a completed demonstration. A task performed entirely with accessible cloud resources does not establish that a workflow involving local applications will remain available overnight. The location of the required work matters.

Consider a hypothetical briefing assembled from an online planning document and a presentation stored on a laptop. The assistant might continue researching the planning material after the laptop disconnects. It would still lack access to the local presentation until the connection returns.

The useful response would identify the missing source and explain its effect on the result. A polished briefing that silently omits the presentation would create a different problem. The output could appear complete while representing only part of the available evidence.

This makes access status part of editorial judgment. Users need to know whether an assistant inspected the relevant material or merely completed the portion it could reach. A task’s apparent completion should not conceal gaps in its inputs.

Connected information also introduces questions about authority. An assistant might have access to an outdated document, an informal discussion, and an approved plan. Reading all of them does not automatically reveal which should govern its actions.

A person often resolves that uncertainty through relationships and institutional knowledge. They know which manager owns the decision or which document reflects the latest agreement. An assistant needs an explicit basis for making the same distinction.

The pressure therefore extends to the organizations adopting these products. Teams with unclear ownership and conflicting records should expect those weaknesses to affect automated work. Giving an assistant more access does not repair an ambiguous operating process.

The commercial implication is straightforward. An always-on assistant competes partly on its ability to explain its own limits. Users deciding whether to rely on a result need a clear account of what the assistant could access.

OpenAI dots should consequently be evaluated across interrupted and incomplete workflows, alongside successful ones. Recovery from missing information is part of the product’s value. So is recognizing when the missing information prevents a responsible answer.

Anthropic’s Cowork Shows the Broader Pressure on Work Software

OpenAI is pursuing a direction that already extends beyond its own products. Anthropic’s Cowork architecture brings capabilities associated with Claude Code into general knowledge work. Anthropic describes cloud execution, document creation, research, and scheduled tasks within that offering.

The competitive pressure concerns who organizes delegated work, not simply who generates the best response.

The comparison establishes a shared direction without establishing equivalent performance. Both companies describe systems that carry out work beyond an immediate exchange. Their documentation alone does not determine which product handles ambiguous instructions or changing priorities more reliably.

That uncertainty matters for software buyers. A feature comparison can show whether a product supports a particular connection or recurring task. It tells less about whether the assistant recognizes a consequential exception during that task.

Suppose a recurring report normally follows a predictable structure. A new business condition makes the usual comparison misleading. Completing the report on schedule would satisfy the mechanical instruction while failing its practical purpose.

The relevant competitive advantage would be recognizing the exception and presenting a useful decision to the user. That behavior requires more than access to a scheduling feature. It requires judgment about the relationship between the instruction and its intended outcome.

This creates pressure on established work applications as well. An assistant that coordinates activity across tools can become the place where users assign and review work. Individual applications still hold records, but the assistant increasingly shapes how people encounter them.

For application vendors, that raises a question about visibility. If users interact through an assistant, the application needs ways to expose reliable state and receive authorized changes. Otherwise, the assistant risks acting on an incomplete interpretation of the workflow.

For employers, the immediate issue is responsibility. Someone still needs to own the final report, customer communication, or operational decision. Delegating preparation does not remove the need for an accountable person.

For employees, the benefit depends on supervision cost. An assistant that needs frequent corrections can turn execution into another management task. The comparison that matters is the total effort required to obtain a trustworthy outcome.

This is why a broad claim that personal assistants will replace existing work software remains premature. The documented products establish an attempt to coordinate more work. They do not establish that users will transfer their habits, trust, and daily attention.

OpenAI and Anthropic therefore face a similar adoption question. Can their products make delegation dependable enough that users return with ongoing responsibilities? Occasional impressive results will not answer that question on their own.

Permission Is the Boundary Between Assistance and Representation

The most consequential distinction in a personal assistant is between understanding a situation and changing it. Reading a message, drafting a response, and sending that response involve different levels of authority.

OpenAI’s documentation on action controls says actions undergo review against instructions, permissions, and safeguards. It explicitly separates permission to draft from permission to send. It also says proactive research cannot itself modify connected applications or send messages.

The same documentation warns that custom rules can be misunderstood. It distinguishes pausing the assistant from stopping delegated tasks or canceling schedules. These are product boundaries users need to understand before relying on continued background work.

Useful autonomy requires specific authority, and authority must remain understandable after a task has started.

Consider an assistant asked to prepare a response to a customer complaint. The drafting assignment permits analysis and composition. Sending the message would represent the user to someone else and potentially create a business commitment.

The content might be accurate and the tone appropriate. That would still leave a separate question about authorization. Output quality cannot substitute for permission to communicate.

Ongoing instructions make the boundary more complicated. A user might authorize routine acknowledgments while reserving promises about refunds or delivery dates. The assistant then needs to distinguish a familiar message format from a consequential exception.

This is the practical meaning of granular control. Users need boundaries that match the decisions they make in real work. A broad setting labeled “email access” would not explain which communications can proceed independently.

There is also a burden on interface design. When an assistant asks for approval, the user needs enough context to make that approval meaningful. An unexplained confirmation prompt transfers responsibility without necessarily improving understanding.

A useful request would identify the proposed action, its recipient, and the relevant consequence. It should also make clear whether approval covers only that action or future actions within a defined scope. Otherwise, repeated prompts can obscure the actual decision.

The social boundary extends beyond the owner’s permissions. A recipient has their own expectations about who is participating in a conversation. Authorization from the sender does not guarantee acceptance from the person receiving an automated interaction.

A meeting makes that issue especially visible. Participants may expect immediate judgment, negotiation, and responsibility from another human. An assistant’s ability to summarize the discussion does not establish that everyone accepts it as a representative.

The more promising arrangement is often preparation followed by human participation. An assistant can organize evidence and outline unresolved issues. The owner can then bring judgment and accountability to the conversation.

That division is not fixed for every task or organization. It is a decision that must be made explicitly. Personal assistance becomes harder to trust when the product encourages users to overlook it.

What to Watch as OpenAI dots Takes on Ongoing Responsibilities

The next evidence should come from ordinary use rather than increasingly elaborate demonstrations. Over the coming months, the important signals concern whether the assistant reduces supervision while preserving clear responsibility.

The strongest evidence will show useful continuity, understandable control, and outcomes that remain valuable after review.

First, watch for evaluations of proactive assistance that include both useful findings and unnecessary interruptions. An alert can be factually correct while arriving too late or concerning an issue the user already resolved. Counting alerts alone would miss those differences.

A meaningful evaluation should examine whether the finding changed an outcome. Did it reveal an overlooked dependency, provide enough context, and reach the appropriate person? Evidence of that behavior would strengthen the case for ongoing assistance.

Frequent low-value suggestions would weaken it. Users who must inspect every interruption to discover whether it matters inherit another stream of work. Proactivity needs to earn attention rather than merely consume it.

Second, watch how clearly the products communicate continuing authority. Updates to permission settings matter when users can understand their scope and see why an action proceeded. More settings alone would not establish better control.

The relevant signal is a workflow that remains predictable when circumstances change. If a previously authorized task encounters a new recipient or consequential decision, the assistant should make that boundary visible. Consistent behavior would support broader delegation.

Confusing approvals or unclear stopping behavior would point in the other direction. Users need to know what remains active after they intervene. Confidence depends partly on whether the system’s behavior matches their understanding.

Third, watch for sustained use of ongoing responsibilities across both dots and Cowork. Repeated use would be more informative than isolated examples of impressive output. It would suggest that users find the arrangement worth maintaining.

Even then, usage needs interpretation. A frequently used assistant might still require substantial correction. The stronger signal would combine continued use with evidence that reviewing and repairing its work takes less effort than doing it directly.

These observations should also shape individual buying decisions. A personal assistant does not need to handle every responsibility to become useful. It needs a dependable role within the user’s actual work.

For OpenAI dots, that role starts with the relationship between continuity and accountability. The assistant can maintain context and advance preparation. The person remains responsible for deciding which commitments and relationships deserve direct attention.

Choose an ongoing responsibility you can evaluate from beginning to end. Define the expected result, the information it requires, and the decisions that should return to you. Then examine the full outcome, including time spent reviewing corrections and responding to interruptions. Ask whether the assistant preserved context when priorities changed and whether its actions stayed within your understanding of the assignment. Keep the work where it improves your day, and revise the boundary where supervision becomes burdensome. The useful question is whether this form of delegation leaves you with more attention for the decisions and conversations that require you.