Christina Huang

Christina Huang worked on Platform Experience at OpenAI in 2025 and demonstrated agent building on stage at that year's DevDay. In a subsequent interview with Sherwin Wu, published October 7, she explained AgentKit's design around a difficult transition: taking agents from prototypes into production. She described the suite as packaging lessons from API customers who needed help iterating on prompts, optimizing them and writing evaluations, with tools usable independently or together.[1][2]

Before the DevDay AgentKit launch, Huang's team built Agents SDK and launched it alongside the Responses API in early March 2025. She tentatively recalled adding MCP to the SDK around that month. Her explanation for adopting it was practical: “MCP was kind of like the natural protocol that developers were already using.” This earlier SDK work supplied the code-based route that Agent Builder would later use for exported workflows.[1]

At DevDay, Huang described two paths through Agent Builder. Developers could use it as a playground to model systems, write prompts and test them, then export the result to run in their own systems through Agents SDK, including with other models. Alternatively, they could use hosted deployment and a shared visual canvas where subject matter experts could contribute feedback and help build datasets. Importing code back into the builder and running code within it were ambitions she discussed, rather than capabilities she presented as already available.[1]

Huang described a recurring team discussion about how much workflows should be deterministic and how much should be driven by an LLM. Working with customer-facing engineers, the team modeled existing workflows to identify missing capabilities and commonly needed nodes; she asked developers to provide feedback in the same concrete form. Asked about the builder's common expression language, she explained it as a way to represent variables and conditional statements. She also credited engineers working directly with customers for templates drawn from customer-support and document-discovery playbooks.[1]

Huang distinguished the demonstrated binary approval node—approving or rejecting an action—from the more involved human decisions customers needed. Some customer workflows, she said, involved human interactions over weeks. Exploring those cases more deeply was a development priority; her account did not present the initial approval node as a complete solution for them.[1]

Huang linked prompt optimization to the difficulty of learning each new model's quirks, recalling a substantial optimization launch alongside GPT-5. She wanted evaluations to guide suggested improvements to prompts and agents over time, describing complete integration of that feedback loop as a direction for further work. She also highlighted the DevDay evaluation launch's support for third-party models as a response to developers using multiple models.[1]

Huang said repeated customer implementations of internal agent builders informed the decision to provide shared prompt and template management, allowing companies to concentrate on work specific to their businesses. She applied a similar argument to ChatKit's chat interface: its embeddable iframe could stay “evergreen” as capabilities changed, without requiring developers to rebuild frontend components. She compared this reuse of common interface elements to Stripe's Elements and Checkout. Wu credited Tyler, David and Jiwon with building the ChatKit components.[1]

Huang said her team relied heavily on Codex pull-request reviews. In the October 2025 interview, she estimated that work on the visual Agent Builder had begun “probably less than two months ago” and said it would not have been possible without Codex, while emphasizing that generated work was checked and edited. She recalled hearing of teammates starting several Codex tasks during a roughly 30-minute bus commute. She said the tasks could help teammates orient themselves in the codebase even if they chose not to use the resulting pull request and instead wrote the code themselves.[1]