What comes after conversational bots as the first era and autonomous agents as the second? Tara Seshan, who leads product for OpenAI Codex and ChatGPT Work, envisions a third phase: an AI colleague that works persistently and collaborates with you iteratively over time.

Speaking on the Lenny’s Podcast on August 30, Seshan explained that AI is transforming knowledge work from “rowing” to “steering.” As agents take over an increasing share of execution, humans will focus on setting direction, evaluating results, and adjusting course. Once execution becomes a commodity, the differentiator for individuals and companies will shift from capability to judgment, taste, and ambition.

This perspective is not an official product roadmap from OpenAI but a front-line leader’s assessment of the future of work. It raises key questions: why are coding agents moving into knowledge work, why can’t enterprises simply purchase a chat interface, and why does the role of managers become more critical as AI grows more capable?

Distinguishing AI Colleagues from Chatbots

Seshan categorizes AI products into three stages: chat, agent collaboration, and the emerging stage of a “continuously working AI colleague.” Chatbots handle a single query with a single response. Agents execute sequentially toward a goal. An AI colleague, however, functions like a team member: it works independently for a period, receives feedback from a human, then continues. Both parties review in-progress work and sync at their own pace. Seshan calls this “letting the agent cook,” where humans don’t dictate every step but provide direction at a higher level of abstraction.

This relationship must also evolve from one-person-one-agent to multi-person collaboration. Today, some internal teams at OpenAI share screenshots of Codex threads in Slack to explain how a number was derived. But screenshots are not a natural interface for collaboration. The future requires enabling my agent’s analysis to be reviewed by a colleague’s agent, integrating multiple humans and multiple agents into a single workflow rather than isolated personal automations.

Sustained work, however, depends not only on smarter models. Seshan used a simple analogy: locking a new employee in a room without access to Google Docs, Slack, or company databases renders even the smartest person useless. A cloud-based agent needs enterprise data, third-party systems, cloud infrastructure, and reliability guarantees. An “AI colleague” is first and foremost an environment equipped with appropriate context, permissions, and tools.

From Rowing to Steering: Elevating Human Roles

“Rowing” involves completing specific tasks personally; “steering” means deciding where the ship goes and adjusting direction based on feedback. Seshan believes agents will handle increasingly longer loops, pushing human control points from completing a line of code up to deliverables, objectives, and even higher levels. But steering is not simply writing a one-line goal and waiting. Data can inform direction, yet many crucial decisions rely on human judgment, intuition, and proactive vision. It’s not that another path is unviable; it’s that people want products and the world to become a certain way.

She compares software to film rather than real estate. Real estate investments typically correlate with greater asset value; high-budget films, however, do not guarantee success. Software similarly requires authorial expression, trade-offs, and taste. When every company can access similar models and agents, tools become commoditized, making human conviction the distinguishing factor.

This shift impacts enterprises beyond reducing manual tasks; it demands changes in performance metrics and organizational structure. Previously, rewards favored those who mastered execution. In the future, companies must identify individuals who ask better questions, set higher-quality objectives, detect when agents go off track, and take responsibility for final outcomes.

When Execution Becomes Cheap, Ambition Becomes the New Constraint

Seshan observes that the most effective AI users do not merely automate repetitive tasks; they expand what they attempt. Previously, someone skilled in product, design, and engineering simultaneously was a rare “unicorn.” Now, an individual can rapidly generate designs, build prototypes, and model pricing and scenarios. Work that once exceeded an individual’s capabilities is now within reach.

Consequently, the limiting factor shifts from execution ability to “thinking bigger.” She advises managers to continuously ask: Is there a more ambitious version? Can we try this faster? Can we scale this tenfold? In her view, elevating a team’s ambition is becoming a core responsibility for product managers.

Internal phrases at OpenAI reflect this culture: “Is this maximally accelerated?” and “Are you mainlining it yet?” The first pushes for speed; the second demands that product teams use their own products intensively, closing the feedback loop with real work.

Product planning is even more aggressive. Seshan argues that designing for today’s model capabilities fails, as does designing for a hypothetical future a year out. OpenAI aims to target model capabilities about two to three months ahead. Product teams must stay closely aligned with research roadmaps while “getting out of the way” to avoid constraining new capabilities with outdated product structures.

This approach suits a frontier model lab but cannot be directly replicated by ordinary enterprises. The two-to-three-month window relies on insider knowledge between research and product teams; “ambition as a bottleneck” is a vendor leader’s judgment, not a quantitatively verified universal law. Enterprises would be better served by avoiding heavy investment in temporary model limitations and instead focusing resources on internal processes, data, permissions, and business acceptance criteria.

Why ChatGPT Work Conceals Codex

The interview also shed light on OpenAI’s product strategy for pushing coding agents into knowledge work. According to Seshan, ChatGPT’s Work mode leverages Codex’s execution capabilities but strips away coding-specific interfaces like the work tree. Users can build complex financial models and other knowledge tasks. If the same task is posed in Codex, the capability is not weaker; the difference lies in what interfaces and technical details are shown during execution.

OpenAI’s ideal state is that users no longer distinguish between Chat, Work, Codex, models, and harnesses. Users simply describe the task, and the system automatically selects the appropriate model and execution framework. The current multiple entry points are a transitional state as the product migrates from chat to agents.

The real difficulty is that knowledge work cannot adopt coding agents’ acceptance methods. Whether code passes tests can often be judged from output. A strategic report or financial analysis, even if seemingly complete, cannot be deemed “90% correct” based on the final file alone. Users need to see the process, inputs, citations, and intermediate work to understand how conclusions were formed.

Thus, knowledge-work agents must not merely deliver answers; they must involve the human in the journey of forming those answers. Users should see which sources were used, whether citations support conclusions, how key assumptions vary, and where human judgment is required. For B2B products, traceability, process collaboration, and business acceptance are not add-ons; they are prerequisites for coding agents to enter core enterprise workflows.

Enterprises Need More Than an Account

A persistently working AI colleague entering the enterprise requires at least four categories of support. First is context: Can it understand documents, emails, meetings, and business data within permission boundaries, rather than requiring employees to re-upload background every time? Second is action capability: Which systems it can call, whose identity it operates under, and how rollbacks work after write failures. Third is collaboration mechanisms: When do humans check in, when does the agent pause, and how do other employees and agents take over? Finally, accountability: Who validates output quality, and who approves high-risk decisions?

This reshapes the competitive landscape for B2B AI vendors. Models provide general capability, but the truly defensible value lies in connectors, identity and permissions, enterprise context, long-task reliability, process evidence, and team collaboration. An agent that can demo anything is not necessarily more valuable than one that can stably integrate into a specific workflow, withstand audits, and allow human handover.

When selecting vendors, enterprises should not only test how impressive a single output looks. A more effective method is to select real historical tasks and run the product under actual permissions, data conflicts, and anomalous conditions, tracking end-to-end completion rates, human intervention, citation traceability, failure recovery, and cost per qualified task. The value of an AI colleague is not that it never requires humans, but that humans can control a larger scope of work with fewer actions.

What Cannot Be Outsourced: Writing for Thinking

Even as agents take over execution, Seshan reserves one clearly human domain for herself: writing for thinking. She categorizes writing into two types. Weekly reports, status updates, and format conversions belong to “writing for reporting” and can largely be delegated to models. Product direction, strategic judgments, and controversial opinions belong to “writing for thinking,” which she insists on drafting herself. Outlining, writing, revising, and defending against critique is itself the process of clarifying one’s thoughts.

She summarizes her approach as: “Start it yourself, end it yourself.” AI can assist in the middle—research, data supplementation, or challenging viewpoints—but it does not generate her first draft of conclusions. Meanwhile, internal sharing at OpenAI is shifting from “documents” to “prototypes and outcomes”—mocks, not docs; prototypes, not docs. Long documents are now trivially generated and no longer inherently signify depth of thought; interactive prototypes and experimental results drive decisions more effectively.

This captures the dual nature of work in the AI era: execution is delegated to agents, but judgment must not atrophy; reporting can be automated, but the process of forming opinions must be preserved. Otherwise, companies may end up with more documents but fewer people truly capable of steering.

The Real Test of the Third Era Is Organizational

The “continuously working AI colleague” remains an evolutionary direction described by Seshan, not a mature, delivered, unified product. ChatGPT Work, Codex, and multi-agent collaboration are still in the midst of product integration and interaction exploration.

But the direction is clear: AI products are moving from answering questions to executing tasks, and then to participating in work over the long term. Model capability determines how fast it can row; enterprise context and infrastructure determine whether it can board the ship; human judgment, accountability, and ambition determine where the ship sails.

The key to AI’s third era is not how many agents a company owns, but whether it can establish a new mode of collaboration: letting AI work continuously, while humans truly steer.