Overview of AI’s third era: the rise of persistent AI coworkers with Tara Seshan
In this conversation, Lenny Rachitsky talks with Tara Seshan, Product Lead for ChatGPT Work and Codex at OpenAI, about how AI products are evolving from chat, to agents, and soon to persistent AI coworkers. Tara shares what it’s like building at OpenAI, how product management changes in a frontier lab, why the best teams now operate with tighter empirical loops, and how AI is already reshaping knowledge work, software, and product craft.
The Three Eras of AI Products
Tara frames the evolution of AI products in three stages:
- Era 1: Chat — interacting with a model through conversation
- Era 2: Agents — the model does tasks for you in loops
- Era 3: Persistent coworkers — AI systems that behave more like ongoing teammates, collaborating with you over time
Her view is that the next major shift is not just “ask and answer,” but ongoing collaboration with AI coworkers that can do substantial work independently and then sync back with humans.
What Surprised Tara Most About OpenAI
Tara expected OpenAI to feel like a place with hidden master plans and deep internal strategy. Instead, she found:
- OpenAI is unusually open
- Ideas quickly become public product or messaging
- There is less top-down direction than at many large companies
- Everyone acts like a founder
- Teams are close to the market
- Product development feels highly owned and entrepreneurial
- The pace of feedback is extremely fast
- Product ideas move from research to user touchpoints very quickly
Her biggest adjustment was realizing that OpenAI is not a place where strategy is hoarded internally — it is a place where strategy is often immediately expressed in products users can touch.
How Product Management Changes in a Frontier AI Company
Tara emphasizes that classic PM skills still matter, but the emphasis changes dramatically.
What matters more now
- Being prolific and empirical
- Try things fast
- Test with users quickly
- Learn from reality rather than long theoretical docs
- Defining the sharpest possible hypothesis
- The most important PM skill is identifying the “eigenquestion”
- What is the one thing that determines whether the product works?
- Staying tied to research and model capability
- Product decisions must track where models are headed in the next 2–3 months
- Building for where models are today is often already wrong
What matters less now
- Long strategy memos
- Overly academic reasoning
- Perfectly polished artifacts before learning from users
Her core point: in AI, the job is less about proving you thought deeply and more about proving you can learn quickly.
Steering vs. Rowing: How AI Changes Knowledge Work
Tara thinks the future of work will increasingly look like:
- Humans steering
- Agents rowing
That means humans will:
- Set direction
- Provide high-level feedback
- Make opinionated calls
- Adjust based on what the agent produces
Meanwhile, agents will handle more of the tactical execution.
She also argues that this doesn’t eliminate human judgment. Instead, it elevates the importance of taste, intuition, and direction-setting.
Why Ambition Matters More in the AI Era
A major theme in the episode is that AI tools expand what individuals can do — and therefore raise the bar for ambition.
Tara’s view on ambition
AI lets people:
- Prototype ideas faster
- Create designs and models without needing a specialist
- Explore more possibilities from the same starting point
- Bring more of their internal vision into reality
This means the limiting factor is no longer execution alone. The real question becomes:
- How ambitious are you willing to be?
- What are you trying to make possible?
She believes PMs should actively raise the ambitions of their teams by asking:
- “Could we do this faster?”
- “Could this be 10x bigger?”
- “What’s the more ambitious version of this idea?”
ChatGPT, Codex, and Work Mode
Tara explains the current product structure as a transition state on the way to a simpler future.
Current modes
- ChatGPT / Chat mode
- For conversation and search
- Familiar experience with improved capabilities
- Codex
- More development-oriented
- Great for coding and agentic tasks
- Work mode
- Uses Codex under the hood
- Designed for knowledge work tasks like financial models, analysis, and internal team workflows
The long-term vision
Users should not have to choose between tools or understand harnesses, model limits, or internal product distinctions.
- The ideal future is one box
- You describe the task
- The system chooses the right harness/model automatically
The goal is to make AI feel like a natural collaborator rather than a menu of technical options.
How Knowledge Work Differs from Coding
One of Tara’s most important product insights is that knowledge work is not the same as coding.
Coding
- Output is easier to verify
- Tests can validate whether the task was done correctly
- Success is often observable at the end
Knowledge work
- The final artifact alone is not enough
- You need visibility into:
- process
- citations
- inputs
- reasoning
- intermediate steps
That’s why ChatGPT Work needs to act more like a collaborator with transparent work-in-progress, not just a black box that returns a finished answer.
Product Craft: Writing, Prototypes, and the End of Docs-as-Proof
Tara distinguishes between two kinds of writing:
Writing as thinking
- Used to develop ideas
- She still does this manually
- Helps sharpen her reasoning and avoid “AI brain rot”
Writing as reporting
- Status updates
- Launch plans
- Summaries and translations
- This she is happy to automate
Her broader point: in this era, prototypes and real artifacts matter more than polished docs. A doc no longer proves you’ve thought something through; a working prototype or experiment is far more persuasive.
What She Learned at Sutter Hill
Tara’s time as an EIR at Sutter Hill gave her a few lasting lessons:
- Product-market fit is not just luck
- There is a repeatable playbook
- Product marketing fit matters earlier than she expected
- Positioning and narrative can be tested before the product exists
- Pitching is a real discipline
- Talk to many people
- Refine the story
- Validate the message before locking the product shape
She came away with a stronger appreciation for messaging, positioning, and enterprise narrative as foundational to product success.
Practical AI Uses Tara Recommends
Tara shared several concrete ways she uses AI that others may not think of immediately:
1. Build sites
She uses Codex/Work to create shareable sites for:
- team games
- trip planning
- dashboards
- presentation artifacts
- personal tools
She sees this as a form of personal software — prompting software into existence for a specific need.
2. Use /visualize
She recommends using a visualization command to turn activity or usage data into a useful story or chart.
3. Let Work mode handle long-running tasks
Especially on mobile, she likes starting tasks and letting them continue in the cloud while she moves on.
Human Skills That Still Matter Most
Tara is clear that humans remain essential in a few areas:
- Accountability
- Someone must own the outcome
- Expression
- What we choose to build still reflects taste and authorship
- Care and collaboration
- Supporting teammates and elevating each other’s ambitions is deeply human
- Opinionated direction
- AI can execute, but humans still decide what direction matters
Her answer is not “humans become irrelevant,” but rather that human judgment becomes more visible and more valuable.
Lightning Round Highlights
Books Tara recommends
- Barbarian Days by William Finnegan
- A favorite for its meditation on dedication and passion
- Anna Karenina
- A book she says reveals new layers as you grow older
Movies / TV
- The Odyssey — she reads it as a meditation on AI and its societal consequences
- Rashomon — admired for its narrative structure and creative constraints
Favorite AI products
- OpenAI tools, especially ChatGPT, Codex, Work, Sites, and Visualize
- A friend’s private social app called Gats
- A podcast-generation app made by a friend
Life motto
She quotes Toni Morrison’s essay The Work You Do, the Person You Are:
- Do the work well for yourself, not just the boss
- You make the job; it doesn’t make you
- Your real life is with your family
- You are not the work you do
Key Takeaways
- AI product development is moving from chat to agents to persistent coworkers
- Product teams should optimize for short feedback loops, not long theoretical planning
- The most valuable PM skill is sharpening the core hypothesis and testing fast
- AI expands capability, which means ambition becomes the new differentiator
- Knowledge work requires transparency into process, not just outputs
- Humans still matter most for accountability, taste, and collaboration
Final Thought
Tara’s core message is that the AI era is not just about better tools — it’s about rethinking how work happens, how teams collaborate, and how ambitious people allow themselves to become. The future she describes is one where humans steer more, agents row more, and the best products feel less like software and more like durable, collaborative coworkers.