Creating Agents that Co-create from OpenAI

The AI Engineer Summit concluded a few weeks ago in New York. If you missed it, don't worry - I'm here to share key insights on AI agents that can benefit your enterprise!

Karina Nguyen from OpenAI shared valuable experiences on designing delightful AI products. If you're an enterprise leader who has already implemented AI prototypes, these lessons will be particularly helpful.

Design Decisions for AI Products

AI is still relatively new, so we're all seeking design decisions that best fit our tools and use cases. Karina shared insights that influenced several of her products - use these as inspiration!

1. Unfamiliar capability, familiar form. Language models processing 100k tokens in their context window was a clear breakthrough. But how should this capability be implemented? A simple, familiar use case was having language models process uploaded documents and answer questions based on that information. Users love uploading financial reports and asking questions about them.What this means for you: Instead of reinventing how agentic workflows can transform your business, start small to make AI approachable for users and employees. Look for familiar tools and explore integrating AI into those systems.
2. AI helping with async tasks. The beauty of language models is their ability to ingest huge amounts of information. With reasoning models emerging, they can now ponder that information and generate their own reports and ideas. OpenAI's Deep Research exemplifies this approach.What this means for you: If you're wondering which tasks suit AI implementation, note the async tasks in your day - those you typically delegate that require significant information processing. AI can create strong first drafts, similar to reports generated by Deep Research.

Formalizing This - A New Type of Product Research Cycle

With extensive experience creating AI products, Karina has developed a product research cycle tailored for AI agents. This focuses on experimentation and research - most relevant for those with dedicated resources or where AI is a core product.

The cycle:

• New task creation - what new things can AI help you accomplish?
• More complex RL environments - train models on diverse tool use (image editing, etc.)
• Leverage in-context learning - enable models to learn new tool usage
• Synthetic data via distillation from stronger reasoning models - run smaller models at lower costs while maintaining performance
• Invent new model behaviors and interactions - where human creativity shines
• Utilize user feedback

Does Karina's experience reflect your experience of implementing AI into your enterprise? Drop a comment below!