Persistent AI Coworkers Mark OpenAI’s Proposed Third Era

OpenAI has outlined a third era for its products, despite persistent AI coworkers remaining more vision than finished product. The claim reached many readers through a google news headline, but its implications extend beyond another workplace feature.

Tara Seshan, OpenAI’s product lead for Codex and ChatGPT Work, described three stages during an August 30 appearance on Lenny’s Podcast. Chat came first. Task-completing agents followed. Persistent coworkers represent the proposed next stage.

That progression creates the central tension. OpenAI wants AI to become a continuing participant in work, yet reliability, authorization, memory, and team coordination remain unresolved. Anthropic, Microsoft, and other enterprise vendors are pursuing related ideas, making this a contest over the future workplace interface.

What OpenAI Actually Said About Its Third Era

OpenAI’s third-era framing describes an ongoing working relationship, not a newly released autonomous employee.

Seshan’s comments originated during an interview about product development, Codex, ChatGPT Work, and the changing role of knowledge workers. The podcast episode ran for about 82 minutes and appeared on August 30, 2026.

She divided the evolution of AI products into three periods. The first centered on chat, where a user asks a question and receives a response. The second introduced agents that pursue goals and complete multi-step tasks.

The proposed third period involves persistent AI coworkers. These systems would remain connected to projects, retain relevant working context, and collaborate across longer periods. They would not start from an empty prompt whenever a user returned.

Seshan presented this stage as something that “might come soon,” not as an established product category with settled specifications. That distinction matters because the initial report compresses a wide-ranging discussion into a definitive headline.

A google news reader could reasonably interpret that headline as a formal OpenAI announcement. The underlying source supports a more cautious conclusion. A senior product leader described the direction she expects AI interfaces to take.

The word “persistent” carries several possible meanings. It can refer to memory across sessions, continuing execution, recurring scheduled work, or access to shared project materials. It can also describe an agent that regularly checks progress with its human collaborators.

OpenAI has not published one technical definition covering every part of that vision. The interview instead connects several existing capabilities with an unfinished product direction.

ChatGPT Work already handles longer assignments across files and connected applications. Codex can divide technical work among parallel agents. Projects and synchronized conversations preserve more context across devices and sessions.

Those components point toward persistence without fully creating an AI colleague. A genuine coworker must understand changing goals, recognize organizational boundaries, and explain what it did. It must also know when to stop.

The proposed interface changes the unit of interaction. Chat starts with a message. An agent starts with a delegated task. A persistent coworker would start with an ongoing responsibility and maintain a working relationship around it.

Consider a product launch. A chat assistant might summarize customer interviews. An agent might analyze those interviews and draft a report. A persistent coworker would track the launch, update the analysis, flag contradictions, and join later planning cycles.

That continuing role makes the idea more consequential than another chatbot upgrade. It also introduces more ways for incomplete context or excessive access to produce errors.

The headline therefore captures a real OpenAI thesis, but not a completed transition. The third era is a destination that OpenAI is building toward through Work, Codex, apps, memory, and collaborative interfaces.

Why Persistent AI Coworkers Matter Now

The shift matters because OpenAI is moving from answering workers to participating in workflows that produce finished business artifacts.

OpenAI introduced ChatGPT Work on July 9, 2026. The company says the agent can operate across applications and files, remain with a project for hours, and create completed materials.

Those outputs include documents, spreadsheets, presentations, reports, and websites. Users can follow progress, answer questions, redirect the work, and approve important actions.

This is different from generating text inside a conversation. The agent must gather information, decide which steps to take, use external software, and maintain enough state to complete a longer assignment.

OpenAI’s Work announcement positions the product as a partner for ambitious tasks. That language starts to resemble the persistent-coworker thesis, although Work still operates around assignments initiated by users.

Seshan also described a narrow product-development horizon. Her team tries to build for models expected within two or three months. Designing only for current capabilities produces an outdated interface, while predicting a year ahead becomes speculative.

That approach explains why OpenAI is discussing the third era before every supporting capability feels mature. Product teams must prepare interaction models for abilities that are still arriving.

The timing also reflects changing usage inside OpenAI. The company says Codex moved from a primarily engineering tool into work performed by legal, finance, recruiting, and other departments.

In May 2026, more than 70 percent of users reportedly assigned Codex work that would take a person over one hour. By June, users at the 99th percentile generated over 60 hours of daily agent activity across parallel tasks.

These figures come from OpenAI’s own analysis, so they should not be treated as independent measurements. They still reveal which behavior the company considers strategically important.

The Codex usage study says nondeveloper adoption grew faster than developer adoption after August 2025. Research usage inside OpenAI also rose sharply during the measured period.

Longer assignments create pressure for continuity. Users cannot efficiently supervise several agents if every session requires reconstructing the project, relevant decisions, and organizational constraints.

Persistence promises to reduce that overhead. An agent that remembers approved goals, recent work, and open questions can resume with less explanation. Shared context also makes delegation easier across a team.

This resembles the purpose of a personal knowledge base. Stored information becomes useful when a system can retrieve the right context during active work.

However, organizational memory is harder than personal retrieval. Companies contain conflicting documents, changing policies, access restrictions, and informal decisions. A coworker must distinguish current guidance from obsolete material.

Teams also need to understand where the agent obtained its context. A polished output is less valuable when reviewers cannot trace its assumptions or determine which source controlled a decision.

Persistent agents therefore pressure several existing software categories. Project-management systems must accommodate machine participants. Knowledge platforms must serve live context. Identity systems must represent delegated authority.

Business applications also face a new interface question. They can remain destinations where people perform work, or become services that agents operate under supervision.

This is why the third-era framing deserves attention. OpenAI is not merely predicting better answers. It is proposing that the workplace interface will organize people, models, data, and tools around continuing goals.

Google News Captured the Claim, but Not the Product Gap

The google news framing makes the transition sound complete, while OpenAI’s own language describes a direction with major missing pieces.

News aggregation rewards concise declarations. “Persistent AI coworkers are OpenAI’s third era” travels more easily than a qualified discussion about future interfaces, product experiments, and model planning horizons.

That compression is useful for discovery. It is less useful for understanding what buyers can deploy today.

ChatGPT Work can perform lengthy assignments, but an assignment is not the same as an enduring role. Codex can run parallel agents, but coordinated execution does not automatically create shared organizational understanding.

Memory also remains narrower than human continuity. A system can store conversation history or project files without understanding which past decisions remain binding. Retrieval and judgment are separate problems.

The difference becomes clear in recurring work. Imagine an agent preparing a weekly product update from interviews, support messages, analytics, and engineering notes.

The first run requires access, instructions, output standards, and definitions for important metrics. A persistent coworker should reuse that setup during later runs without freezing outdated assumptions.

It should notice when the launch plan changes. It should ask whether a new metric replaces the previous one. It should preserve earlier evidence while avoiding conclusions that no longer apply.

It must also coordinate with multiple people. A product manager, designer, and engineer can disagree about priorities. Shared access does not tell an agent whose instruction controls the final artifact.

Seshan described the future interface as multiplayer. Groups of people would steer groups of agents around shared information and work products. That model goes beyond the familiar pattern of one user directing one assistant.

Multiplayer work introduces ownership questions. Someone must know which agent created an artifact, which human approved it, and which system supplied the underlying data.

Conflicts become inevitable when agents receive overlapping responsibilities. Two agents might edit the same presentation, contact the same customer, or interpret a policy differently.

Software teams already handle similar coordination through version control, access rules, reviews, and audit logs. Persistent coworkers require equivalent controls across less structured knowledge work.

This helps explain Seshan’s distinction between “steering” and “rowing.” Agents increasingly perform the execution, while people choose direction, evaluate tradeoffs, and make opinionated decisions.

The metaphor is compelling, but steering can demand more attention when execution is difficult to inspect. A manager supervising several opaque agents might spend less time producing work and more time verifying it.

That risk is especially relevant when AI creates finished materials. Errors can move from an answer into a spreadsheet, presentation, customer message, or operational system before someone notices them.

A useful persistent coworker therefore needs more than longer runtime. It needs durable context, explicit authority, visible provenance, controllable actions, and reliable escalation.

OpenAI has pieces of that stack. ChatGPT apps connect external sources. Work supports approvals. Codex provides structured task execution. Enterprise controls restrict which resources users can access.

The product gap lies in combining those pieces into a dependable relationship that lasts across projects, people, and changing conditions.

Readers arriving through google news should treat the third-era label as a strategic map. It is not evidence that a general-purpose digital colleague has already arrived.

OpenAI Is Competing to Own the Coworker Interface

The primary contest is between persistent AI collaboration and the current model of isolated, task-based agents.

OpenAI’s closest opponent is not one company. It is the prevailing interaction pattern where users open a tool, assign one task, inspect the result, and begin again.

That pattern limits autonomy, but it contains failures. The agent receives temporary context, operates within a bounded session, and usually stops after delivering an answer or artifact.

Persistence removes some of those boundaries. The agent gains more time, more context, and potentially more access. Those additions can improve usefulness while increasing the cost of mistakes.

OpenAI wants to make persistence manageable through products that connect chat, knowledge work, and coding. Seshan described Chat, Work, and Codex as distinct modes for different kinds of activity.

Chat remains useful for search, analysis, and conversation. Work handles broader knowledge assignments. Codex focuses on software development and technical execution.

A persistent coworker could eventually move across those modes. It might discuss a goal in Chat, build an analysis through Work, and ask Codex to create supporting automation.

That integrated path gives OpenAI an important advantage. The company can connect a widely used conversational interface with agent execution, developer tooling, and workplace data.

However, Anthropic is pursuing similar long-horizon behavior. Its Claude products emphasize memory across sessions, extended work, coding agents, and enterprise workflows.

Anthropic says Claude Opus 4.7 can carry context across sessions and manage multi-day projects. Like OpenAI’s claims, that description comes from the vendor and requires evaluation in specific environments.

Microsoft also approaches the market through applications, enterprise identity, and Copilot agents. Its distribution across Microsoft 365 gives it immediate access to documents, meetings, email, and organizational permissions.

That competition is not simply about model scores. The company that becomes the workplace’s coordination layer can mediate how agents see data, request approval, and produce artifacts.

Trust may matter more than raw autonomy. Enterprise buyers need controls that reflect existing identities, retention policies, regulatory duties, and separation between departments.

Context quality matters as well. A model with extensive access can still select the wrong document or overlook a decision made during an unrecorded meeting.

This creates an opening for knowledge systems that organize personal and team information before an agent acts. A dependable AI workflow needs clear sources and review points, not just a longer prompt.

OpenAI’s product strategy also challenges traditional software pricing and usage models, even without discussing specific commercial terms. Software has often sold seats to human users. Persistent agents consume resources continuously and may perform work across several applications.

Vendors must decide whether agents count as users, integrations, automations, or a new category. Customers must determine who owns their outputs and which department pays for the underlying activity.

There is also a cultural contest. Task-based agents fit established delegation habits. Persistent coworkers ask teams to treat software as a participant with continuing responsibilities.

That shift requires new management practices. Teams need to assign scope, review outputs, record decisions, and remove access when responsibilities change.

OpenAI can supply the interface, but employers must redesign the operating model around it. Adoption will stall if organizations cannot explain who remains accountable.

The winning approach will make persistence feel controllable. Users must be able to inspect ongoing work, correct assumptions, limit authority, and recover from errors without restarting an entire project.

That is a higher bar than producing an impressive demonstration. It requires reliability across ordinary, messy, and adversarial workplace conditions.

Persistence Expands Both Capability and Risk

Every feature that makes an AI coworker more useful also gives mistakes and attacks more time to spread.

A chat error usually affects one response. A persistent-agent error can shape later decisions because the system may store, retrieve, and build upon the flawed conclusion.

Incorrect memory is one concern. An agent might preserve a preliminary decision as final, retain outdated customer information, or associate an instruction with the wrong project.

Access creates another problem. A coworker that prepares reports across email, documents, calendars, and internal databases needs broad visibility. That visibility can expose sensitive material when permissions are poorly scoped.

Action rights raise the stakes further. Reading a customer record is different from editing it. Drafting an email is different from sending it. Preparing code is different from deploying it.

Persistent coworkers also consume untrusted content. Emails, websites, documents, and messages can contain text designed to manipulate an agent through indirect prompt injection.

NIST describes agent hijacking as an attack where malicious instructions appear inside data that an agent processes. The system may then perform an unintended action while believing it is completing the user’s task.

In one NIST evaluation, repeated attempts raised the average success rate across selected attacks from 57 percent to 80 percent. Those results came from a specific experimental setup, not every deployed agent.

The finding still exposes a persistence problem. A long-running system encounters more content and offers attackers more opportunities than a short, isolated conversation.

NIST’s agent security work highlights indirect prompt injection, data poisoning, specification gaming, and harmful actions without adversarial input. The agency is developing guidance for single-agent and multi-agent systems.

Identity and authorization become central defenses. Each agent needs an identifiable role, limited permissions, and records showing which actions it took under whose authority.

A persistent coworker should not inherit every permission held by its supervisor. It should receive only the minimum access required for its assigned responsibility.

Approval checkpoints also need context. A vague confirmation box cannot protect users if it hides the proposed action, destination, data involved, and downstream effects.

OpenAI says ChatGPT Work allows users to approve important actions. That is a meaningful control, but the company has not established that every consequential workplace action will be classified correctly.

Human review has limits too. Frequent low-quality approval requests create fatigue. Users may approve actions automatically, especially when an agent has performed many earlier steps correctly.

Long workflows complicate evaluation. A final document might look accurate while depending on an incorrect assumption introduced hours earlier. Reviewers need provenance for intermediate decisions, not only a finished artifact.

Persistence also raises privacy questions. Useful memory requires retention, but unnecessary retention increases exposure. Teams need controls for deletion, correction, expiration, and separation between projects.

The multiplayer model adds another layer. One agent’s output can become another agent’s input, creating a chain where the original error becomes harder to locate.

Organizations should therefore evaluate persistent coworkers as systems, not as models. Tests must cover connectors, permissions, memory, human approvals, auditability, and recovery behavior.

OpenAI has described security and governance controls for enterprise deployments. However, the broad third-era vision has not yet received independent validation across typical organizations.

The cautious conclusion is not that persistent agents are unusable. It is that persistence changes the risk model as much as the productivity model.

The google news headline emphasizes the destination. Enterprise adoption will depend on evidence that the route includes reliable controls, measurable failure rates, and clear accountability.

Three Signals Will Show Whether the Third Era Is Arriving

OpenAI’s third-era thesis becomes credible only when persistence, multiplayer coordination, and controlled autonomy appear in real deployments.

The first signal is a product that maintains project state across days without forcing users to rebuild context. ChatGPT Work already continues long tasks, but durable coworker behavior requires more.

Watch for explicit controls over what the system remembers, which sources remain authoritative, and when stored context expires. Corrections should propagate without erasing the history needed for accountability.

If OpenAI ships these controls with visible provenance, the third-era claim gains support. If persistence remains conversation history plus broad retrieval, the distinction from current assistants stays weak.

The second signal is a genuine multiplayer interface. Seshan described groups of people steering groups of agents, yet most commercial interactions still center on one person and one active agent.

A convincing implementation would show shared ownership, role-based instructions, conflict resolution, agent identities, and review queues. It would also prevent two agents from silently overwriting each other’s work.

That interface must work outside controlled demonstrations. Teams should be able to see which person changed direction, which agent acted, and which artifact represents the accepted result.

Successful multiplayer coordination would strengthen OpenAI’s argument that AI is becoming a coworker. Failure would suggest that agents remain personal tools attached to individual accounts.

The third signal is independent evidence about reliability and security. Vendor case studies show what products can accomplish, but buyers need failure rates under realistic conditions.

Look for evaluations involving prompt injection, stale memory, conflicting instructions, revoked permissions, and recovery after interrupted workflows. Audits should measure consequential errors, not only task completion.

Regulatory and standards activity will also influence adoption. NIST is examining agent identity, authorization, monitoring, and security controls. Clear standards would help customers compare systems using common expectations.

Competitive responses matter within all three signals. Anthropic and Microsoft can weaken OpenAI’s position by offering more dependable memory, governance, or coordination before OpenAI completes its integrated vision.

OpenAI’s advantage is the breadth of its product surface. Its challenge is turning Chat, Work, Codex, apps, and enterprise controls into one coherent relationship.

For knowledge workers, the practical question is not whether an AI deserves the title “coworker.” The better question is which responsibilities it can hold without making supervision harder than execution.

Start with a recurring, reviewable workflow. Define its sources, permissions, approval points, and owner. Then measure correction time, missed context, and consequential errors across repeated runs.

That process tests the substance behind the google news claim. Persistent AI coworkers will represent a third era only when they can preserve context without preserving mistakes.

They must coordinate without obscuring responsibility and act without exceeding their authority. Until those conditions are demonstrated, OpenAI’s third era remains a serious product thesis rather than a completed workplace transition.