Sherwin Wu

As of February 2026, Sherwin Wu led engineering for OpenAI’s API platform. His work includes collective credit for the 2023 function-calling launch and public technical explanations of how developers connect models to tools, evaluate agents and monitor API reliability.[6][3][4][1]

At OpenAI DevDay in October 2025, Wu described developers and third parties as essential to bringing the benefits of OpenAI’s technology to the world. He viewed distribution through ChatGPT apps as an extension of the API’s purpose. Recalling feedback about earlier plugins, he said companies had too little control over how their integrations were used; the newer approach let developers steer the experience and preserve their brands.[1]

Wu credited Anthropic with treating MCP as an open protocol, including an open process for making changes. He named Nick Cooper as OpenAI’s participant on its steering committee and said the protocol’s generality and simplicity made integration easier. He also expressed interest in portable agent workflows and standardized stateful APIs, so developers would not need separate integrations for every model provider; these were aspirations rather than a completed interoperability standard.[1]

He supported keeping both first-party sync connectors and third-party MCP connectors. Sync connectors require OpenAI to maintain substantial state, but having the data enables techniques such as reranking to improve results. MCP connectors reach a longer tail of services, while leaving model performance partly dependent on the API shape chosen by each connector’s creator.[1]

Wu argued that visual workflows become useful as customers’ agent configurations grow large: a canvas can make a complex flow easier to understand than a long SDK file. In discussing AgentKit with Christina Huang, who performed the Agent Builder stage demo, he credited Tyler, David and Jiwon with building ChatKit’s polished, consumer-grade components.[1][2]

For agent evaluation, Wu distinguished a launched capability from the work still ahead. In October 2025, developers could bring Agents SDK traces into the evaluation product and grade an agent’s overall trajectory. Evaluating and optimizing individual parts of a trace remained roadmap work. He emphasized the difficulty of judging a long task correctly, rather than treating it as a simple prompt-and-completion pair.[1]

Wu’s view was that prompting had become more important as models improved. He favored prompt-based optimization partly because serving many separate fine-tuned model snapshots was operationally difficult. He praised GEPA research, while explicitly saying OpenAI’s product was not then using GEPA optimization. He also recalled discussing a low-level research fine-tuning library with John Schulman at OpenAI and welcomed Tinker’s release, describing the idea as Schulman’s longstanding project.[1]

Wu described changing his own use of Codex after watching interns and new graduates delegate whole features. For his personal projects, he had tended to assign small tasks to a model he treated like an untrusted intern. Younger colleagues encouraged him to “push yourself to like trust the model to do more and more.” He said larger assignments sometimes produced results that surprised him and forced him to reconsider his expectations.[1]

Wu singled out the Service Health Dashboard as a DevDay 2025 launch he cared about despite its lack of stage time. It gave each organization visibility into its own API integration, including token velocity, throughput and response codes. He connected that enthusiasm to reliability work following a difficult outage in December 2024, recalling the customer conversations afterward. He credited the Infer team led by Venkat, working with Jana on his team, for improvements. Five-nines availability was a target, not a result he claimed to have achieved.[1]

Before these agent-platform discussions, Wu was already explaining how developers could connect GPT models to external systems. His 2023 work on function calling provides earlier context for the later discussion of tool connectivity and agent evaluation.[4][3][1]

According to his Sancus Ventures biography, Wu graduated from MIT with a computer science degree, worked on Quora’s News Feed Ranking and Product Teams, and then spent more than five years on Opendoor’s Pricing Team before joining OpenAI. Sancus also lists him as an advisor.[5]

Latent Space’s June 14, 2023 coverage credited Wu, Atty Eleti, Jeff Harris and Logan Kilpatrick with driving the Functions API launch, identifying Wu and Eleti with OpenAI’s Applied AI team. In their joint presentation, published by InfoQ in August 2023, Wu explained the model changes and API mechanics; Eleti demonstrated SQL queries, location and Yelp searches, and pull-request review.[3][4]

Wu explained that OpenAI had fine-tuned GPT-4 and GPT-3.5 Turbo on tool-use and function-calling data. The model receives function descriptions and signatures, then expresses an intent to call a function with constructed arguments. The application executes the function and returns its result to the model, which can produce a user-facing answer. His weather example used get_current_weather: GPT requested weather for Brooklyn, the application supplied a temperature and conditions, and GPT summarized them. Eleti separately emphasized that permissions and validation remained the developer’s responsibility, with model output treated as untrusted client output.[4]

Wu also kept limitations explicit. In the 2023 presentation, he said OpenAI no longer trained on API input data, while distinguishing that policy from enterprise-level per-user data isolation, which he said was not yet present. Asked about multilingual enum errors, he acknowledged that newer models would not be perfect and that the team lacked evaluations for the English/French enum case.[4]