Stop planning for 2027: how OpenAI builds product 90 days at a time | Tara Sesha and Nan Yu (OpenAI)

Key constraints include ensuring products are additive and genuinely useful, passing internal testing for delight and novel use cases, and aligning with model capabilities anticipated 2-3 months in the future, avoiding being too anchored in the present or overly futuristic.

The best approach often maps to human nature, suggesting that managing too many individual agents can be overwhelming. Grouping agents into bundles of activity, perhaps managed by a 'chief of staff' agent, may be more effective. Practical considerations like data access and memory also play a role.

While enterprises may struggle with the breakneck pace of AI advancements, they must adopt new features like agents quickly to avoid being leapfrogged. The principle of building what users *need* rather than what they *say* they need is crucial.

Beyond classic product management skills like user empathy and systems thinking, a relentless drive to iterate, test empirically, and learn from feedback loops at a high clock speed is critical.

OpenAI views ChatGPT as a platform with layers of functionality. They expose hooks for third-party developers and offer fallback tools like 'computer use' to ensure tasks can be completed, aiming for composable and cohesive user experiences.

Onboarding, user understanding of privacy and data usage, and the need for clear product norms have become more critical. The importance of empirical testing and moving faster has also increased, alongside the necessity for product leaders to be highly accessible and maintain direct user relationships.

Key predictions include voice interfaces becoming mainstream due to their natural interaction, 'self-driving' like product experiences that guide users and reduce the 'empty input box' problem, and new hardware form factors that may replace traditional laptops.