DevDay 2025: Apps SDK, Agent Kit, MCP, Codex and why Prompting is More Important than Ever
Timestamps are as accurate as they can be but may be slightly off. We encourage you to listen to the full context.
Podcast Summary
OpenAI's platform team leaders, Sherwin and Christina, join the Latent Space podcast from the DevDay 2024 studio to discuss the major platform launches including Apps SDK, AgentKit, and enhanced developer tools. The conversation covers OpenAI's evolution from 4 million to a sophisticated developer ecosystem serving 6 billion tokens per minute. (01:20)
- Key themes: OpenAI's mission to democratize AGI benefits through developer tools, the evolution from plugins to Apps SDK, and the comprehensive AgentKit platform for building production-ready agents
Speakers
Sherwin Chen
Platform team leader at OpenAI who has been instrumental in API development and developer relations. He has extensive experience working with enterprise customers and was involved in early discussions about research fine-tuning libraries with John Schulman before Schulman's departure to found Thinkr AI.
Christina Kim
Platform team member who led the AgentKit demo at DevDay 2024, successfully building an agent in 8 minutes on stage. She's a former Stripe alum who joined OpenAI and spoke at the Singapore DevDay during her first week. Her team has been working on chat interfaces and agent building tools for over a year.
Key Takeaways
Trust AI Models with Larger Tasks
OpenAI's own engineers discovered that giving AI models bigger, more complex tasks often yields better results than micromanaging with small incremental steps. (39:42) New graduate engineers were observed "full YOLO mode" trusting Codex to write entire features, achieving success rates around 40% compared to the conservative approach of treating AI as an untrusted intern with tiny tasks. This represents a fundamental shift in how developers should approach AI collaboration - moving from safety-first incremental prompting to bold, comprehensive task delegation.
Embrace Iterative Platform Development
OpenAI's approach to building developer tools follows a deliberate pattern of iterative releases and learning from user feedback. (02:22) The evolution from plugins to GPTs to Apps SDK demonstrates how each iteration incorporated lessons learned from previous versions. This methodology allows for real user feedback integration and prevents over-engineering solutions before understanding actual developer needs. Companies building developer platforms should prioritize getting basic versions in users' hands quickly rather than perfecting features in isolation.
Build for Both Internal and External Use Cases
The most successful developer tools serve dual purposes, addressing both internal operational needs and external customer-facing applications. (33:01) OpenAI's AgentKit powers their own customer support at help.openai.com while also enabling companies like Ramp to build sophisticated customer-facing agents. This dual-use approach ensures the tools are battle-tested in production environments while providing clear value propositions for potential customers. Internal usage drives quality and reliability standards high enough for external adoption.
Standardize on Open Protocols When Possible
Rather than building proprietary solutions, OpenAI chose to adopt and contribute to the Model Context Protocol (MCP) developed by Anthropic. (05:27) This decision, made around March 2024, demonstrates the value of embracing industry-wide standards over creating competing protocols. The open governance structure of MCP, including OpenAI having a seat on the steering committee through Nick Cooper, shows how companies can collaborate on foundational technologies while competing on implementation and user experience.
Design for Consumer-Grade Polish from Day One
Even internal or developer-focused tools benefit from consumer-grade design and user experience standards. (33:41) Christina's team built ChatKit with "buttery smooth animations" and "responsive designs" that feel like ChatGPT itself, rather than typical internal tool interfaces. This high design standard attracts external customers who demand polished experiences for customer-facing applications. The investment in visual design and user experience pays dividends in both adoption rates and customer satisfaction.
Statistics & Facts
- OpenAI now serves over 6 billion tokens per minute through their API, surpassing the 6 billion figure mentioned in previous communications. (44:18) This massive scale demonstrates the unprecedented adoption of AI APIs in production applications.
- ChatGPT has grown to 800 million weekly active users and is now the fifth or sixth largest website globally. (04:24) It also holds the number one and number two spots for most downloaded apps on the Apple App Store alongside Sora.
- OpenAI's developer community has grown from 3 million to 4 million developers year-over-year. (01:20) This represents a significant 33% growth in the developer ecosystem using OpenAI's APIs and tools.
