Overview of OpenAI Codex lead on the new shape of product work
In this conversation, Andrew Ambrosino, product and engineering lead for the Codex app at OpenAI, explains how AI is reshaping product development, design, and team structure. The big shift: implementation is no longer the expensive part—judgment, curation, and taste are. He describes a world where many people across the company can build prototypes quickly, which changes product work from planning and execution to selection, synthesis, and steering. The episode also covers how Codex is evolving from a coding tool into a general “home base” for knowledge work, plus how OpenAI thinks about browser use, connectors, automation, and role boundaries in an AI-native workflow.
The New Shape of Product Work
From documentation-heavy to prototype-heavy
- Historically, teams de-risked ideas with:
- PRDs
- research
- design docs
- prototypes
- Now, because models make implementation so cheap, teams can jump directly to working prototypes.
- That creates a new problem:
- people can build too many things too fast
- the bottleneck shifts from building to deciding what is worth building
Why “taste” matters more now
- Ambrosino argues that the key job has become:
- choosing the right medium
- identifying which ideas matter
- curating what to keep, merge, or discard
- “Taste” is not just aesthetics.
- It includes systems thinking
- product judgment
- knowing the right abstraction
- understanding context and business goals
The design process is changing, not disappearing
- He agrees that the old, formal design process is effectively dead in its traditional form.
- But design as a discipline is still critical.
- The main shift is that:
- the process is less sequential
- prototypes can look production-ready much earlier
- teams must be more explicit about what stage something is actually in
How AI Is Changing Product Teams
Roles are blending
- On the Codex team, designers write code, PMs understand technical tradeoffs, and engineers participate in product thinking.
- People are increasingly defined by:
- the average of what they do
- not the strict boundary of their job title
- This creates more overlap between design, product, and engineering.
“Everyone is a builder” has limits
- Ambrosino is skeptical of the idea that product roles should disappear entirely.
- His concern:
- removing roles can erase hard-earned best practices
- not everyone can do everything
- depth and specialization still matter
- He believes the future is less about rigid boundaries and more about fluidity:
- people can move more easily
- but disciplines still exist
Managers won’t disappear
- Even in an agentic world, management remains important.
- The difference is that both ICs and managers are increasingly “managing” something:
- ICs manage agents, workflows, and tasks
- managers coordinate teams and larger systems
What Codex Is Becoming
From coding tool to general work hub
- Codex started as a developer tool and CLI, then became a desktop app.
- Internal usage revealed something bigger:
- people outside engineering also wanted to use it
- teams in finance, legal, comms, and marketing started adopting it
- The product is evolving into a broader knowledge-work surface:
- a place to start work
- automate work
- hand work off to specialized tools
The app is becoming a “home base”
- Ambrosino describes Codex less as a single app and more as:
- a central place to organize tasks
- an interface to other tools
- a launchpad for automation and cross-app workflows
- It may open other tools rather than replace them.
- Example: using Excel via the app rather than rebuilding Excel inside Codex
Codex is used for more than coding
Examples mentioned in the episode include:
- organizing files
- drafting documents
- data analysis
- reading email
- managing Slack noise
- editing videos
- building internal automations
- browser-based workflows and computer use
Product Discovery, Dogfooding, and Workflow Design
Dogfooding shapes the product
- The Codex team uses Codex heavily to build Codex.
- That feedback loop drives rapid improvement:
- if the product can’t do a task internally, that becomes a product problem
- They intentionally use the app in uncomfortable ways to learn where it breaks.
Personal workflows often become product features
- The team watches which workflows users repeatedly invent for themselves.
- If enough people recreate the same setup, it may become:
- a first-class feature
- a native primitive
- a more polished product capability
- Examples discussed:
- daily briefings from Slack
- automations that monitor work streams
- memory-like systems for storing context
- computer use for tasks without connectors
Connectors vs browser vs computer use
- The team is actively figuring out when to use:
- native connectors
- in-app browser
- Chrome extension
- computer use
- The ideal outcome is that users simply ask for what they want and the app figures out the best path.
Planning in an AI-Native World
Long-term planning is fuzzier
- The faster models improve, the less reliable detailed long-range roadmaps become.
- Ambrosino says:
- short-term work needs detail
- long-term plans must stay hazy
- false precision wastes time
Product timing matters as much as product shape
- A product can fail simply because the model is not capable enough yet.
- He gave the example that the Codex app released in February might have failed if it had launched in November.
- Lesson:
- sometimes the same exact product needs to be released multiple times as the model improves
Ambition vs readiness
- Teams should be ambitious and explore things that don’t yet fully work.
- But they must clearly label whether something is:
- a prototype
- an internal artifact
- a testable user-facing feature
- One theme: build now, but don’t assume it is ready just because it exists.
What Great Talent Looks Like Now
High-agency + high-taste people win
- Ambrosino repeatedly emphasizes the importance of people who can:
- move quickly
- think independently
- judge quality well
- take ideas from concept to finished product
- In his view, the most valuable people are increasingly those who can:
- orchestrate work
- steer agents
- know what good looks like
Skill still matters
- He pushes back on the idea that AI makes disciplines irrelevant.
- Example:
- knowing how to use Excel is not the same as being able to do finance work
- Likewise:
- using AI tools well doesn’t eliminate the need for product, design, or engineering skill
Failure, Learning, and Career Perspective
He has failed a lot
- Ambrosino shared that he spent 10–15 years failing through:
- a startup that was eventually sold for parts
- AI tools in a regulated industry that didn’t get traction
- His takeaway:
- success often comes from timing
- skills and market conditions need to line up
- persistence matters
OpenAI is unusually candid internally
- He says the company gives blunt feedback on failed product ideas.
- Large internal message threads can be brutal when something is off.
- That honesty, he suggests, helps improve the external product.
Notable Takeaways
Key ideas worth remembering
- Implementation is cheap; judgment is expensive.
- Product work is shifting from “build and de-risk” to “prototype, compare, and curate.”
- Design is not dead, but its process is changing dramatically.
- Roles are blurring, but specialties still matter.
- The best people in AI-native teams combine:
- agency
- taste
- technical fluency
- adaptability
- Timing relative to model capability can determine whether a product succeeds or fails.
Memorable Lines and Insights
- “The implementation is actually not the expensive part anymore. It’s, dare I say, taste.”
- “We need the tastemakers to guide things from inception to what the product should be.”
- “The average of where you’re working is your role.”
- “You might need to release this thing six different times before it works.”
- “We can build anything—but we still need to choose what matters.”
Practical Implications for Product Teams
What teams should do differently
- Use the right artifact for the job:
- doc for clarity
- prototype for interaction testing
- in-product experiment for real behavior
- Be explicit about stage and intent.
- Expect more overlap between roles.
- Hire for:
- strong judgment
- technical curiosity
- ability to manage AI-assisted workflows
- Treat AI output as a starting point, not proof of readiness.
Closing Thought
The episode’s central message is that AI is not just changing what products can do—it is changing how products are made. The new competitive advantage is less about who can generate code fastest and more about who can best steer, refine, and shape all the possible things AI can produce.