How Rohan Varma Uses Codex for Product Work

OpenAI Codex product manager Rohan Varma discusses how he uses agents for product work with Peter Yang. He describes his own team's practices and uses a dummy environment for demonstrations. The episode is sponsored by Oceans.

Varma says Codex helps him synthesize customer feedback and gather context when he joins a project. He also sees a change in team roles: engineers can build and iterate quickly while product managers focus on key decisions, customer conversations and bringing work to market.

He describes short planning horizons and rapid prototypes. One example is an engineer's early in-app browser prototype, which gave the team something concrete to discuss. Varma's account reflects the Codex team's particular environment and does not imply that every product should skip planning.

He believes many users have not explored the full range of tasks they can delegate. As an example of tool use, he recounts asking an agent to turn a document into a site; when an integration did not provide images, the agent used a browser to retrieve them.

In the demonstration environment, he describes collecting feedback from chat, summarizing it and adding it to a task board. After practicing the process, he creates a recurring automation and specifies what should happen on completion. These examples describe his setup and the capabilities available at the time.

Varma uses image generation to explore interface ideas from a screenshot. He asks for several project-picker variations, then can take a promising direction into a shareable prototype. He also discusses using design guidance and testing skills after updates rather than assuming they remain correct.

For his own workload, he describes generating a small local app that ranks messages requiring a response. If a tool proves useful repeatedly, he may arrange for its data to be refreshed. He also summarizes long conversations and asks for a list of commitments.

Varma says he has many customer meetings and often builds in the evening. He asks an agent to review the week's calendar for meetings that could be consolidated or replaced by a message, creating more focus time.

He describes running several tasks in parallel but checking their results between meetings rather than monitoring every step. One example is asking an agent to watch a pull request, address checks and report when it is ready for human review. The exact permissions and workflow depend on the setup.

He also describes maintaining project overview pages from channel updates so a team can find recent context. His aim is to make information easier to keep current, while the contents still require judgment and verification.

Varma says he uses relatively few personal skills, preferring to see what the base system can do. When he repeats a process, he may turn the successful workflow into a skill or automation. He recommends asking what context was missing when a task fails and then refining the workflow.

Varma estimates that the tools increase his own throughput and argues for trying ambitious tasks. Those are personal observations, not measured productivity guarantees. He closes by describing a shift toward deciding what is worth building and spending time with customers.