Key Takeaways
- OpenAI launched AgentKit and Apps SDK for building and deploying AI agents.
- AgentKit offers a comprehensive suite for agent creation, optimization, and deployment.
- The Apps SDK inverts the app-chatbot paradigm, embedding applications within ChatGPT.
- OpenAI adopted Anthropic's MCP protocol for universal tool connectivity in AgentKit.
Deep Dive
- Sherwin Wu stated OpenAI's core mission is to bring AGI benefits to the world, emphasizing reliance on developers.
- The Apps SDK is a natural extension, building on plugins and GPTs, leveraging ChatGPT's distribution.
- This system is designed to allow developers more control and brand preservation by embedding applications within ChatGPT.
- Christina Cai demonstrated AgentKit, building a customer support agent in under 8 minutes using the visual Agent Builder.
- AgentKit comprises an SDK, Agent Builder, connector registry, ChatKit, and evaluation tools, simplifying agent development.
- The visual Agent Builder includes 'user approval' nodes and was used to deploy a live chat app for DevDay's website.
- It is a foundational, Turing-complete tool integrating human approval workflows for complex decision-making processes.
- Visual agent builders are beneficial for complex workflows, with OpenAI releasing templates for common uses like customer support.
- Discussions include protocol interoperability between different agent builder implementations and standardizing stateful APIs.
- OpenAI's Evals tool now supports third-party models, including those from OpenRouter, for testing and comparison within a unified product.
- Automated prompt optimization is a key area of developer interest, integrated with evals for continuous improvement.
- Prompt optimization is an active research area, contrary to predictions from two years prior, with external contributions noted from Databricks.
- 'Zero gradient fine-tuning' via prompt tweaking is discussed as a practical alternative, offering performance gains without custom model snapshot complexities.
- OpenAI is seeking developer feedback on AgentKit, specifically regarding common expression language versus natural language for conditional blocks.
- Feedback is also desired on the balance between deterministic and LLM-driven workflows, and identifying gaps in existing workflows.
- The team is developing a new language for agent builders, highlighting the common expression language for variables and conditional statements.
- OpenAI supports multiple connector options, including first-party sync connectors and third-party MCP connectors.
- First-party sync connectors store significant state for enhanced quality, while third-party MCP connectors offer flexibility but rely on external API shapes.
- The MCP protocol supports first-party, third-party servers, and open-ended options, facilitating company management of developer access and configurations.
- AgentKit, including ChatKit, supports both internal (data processing) and external (customer service) use cases; help.openai.com already utilizes it.
- ChatKit components were developed with a focus on consumer-grade quality, aiming for a polished user experience comparable to ChatGPT.
- Developers are encouraged to leverage ChatKit's widget builder, which includes AI-assisted creation and a custom domain (widget.studio), to streamline chat integration.
- Newer AI-native developers are observed to fully trust Codex for larger tasks, sometimes succeeding in 30-40% of attempts.
- Codex is used internally for code reviews, with one guest noting it even reviews its own submissions.
- Engineers leverage Codex for tasks like generating code snippets during commutes, assisting with context switching and codebase orientation.