Bill Chen

Bill Chen worked on OpenAI’s Applied AI Startups team in 2025, explaining coding-agent architecture and collaborating with customers on integrations. His contributions include a technical tutorial on reasoning items in the Responses API and work with Cursor on adapting its agent to Codex models. Before that, he cofounded RiskAngle, an AI startup for summarizing medical records.[2][3][5][4]

In Future-Proof Coding Agents, presented with OpenAI colleague Brian Fioca at AI Engineer Code 2025, Chen separated a coding agent into three components: user interface, model and harness. He described the harness as prompts and tools combined in a core agent loop—the software that supplies inputs to the model and handles its outputs. This distinction framed a practical problem: new models can require changes to the agent built around them.[2]

Chen proposed using Codex the agent inside another agent, with the maintained harness becoming an abstraction layer. In his account, this lets product builders reuse prompts and tools maintained alongside the models rather than repeatedly optimize them for each upgrade. Fioca supplied the complementary product argument: reusing that infrastructure gives teams more room to focus on their product’s differentiated value.[2]

Chen also described a route for teams retaining their own agent layer. He and colleagues worked closely with Cursor on its integration with Codex the model, distinguishing it from the complete Codex agent. Cursor aligned its tools with the model’s training distribution and its harness with the open-source Codex CLI implementation. The example made the choice concrete: embed the maintained agent, or keep a custom harness and adapt its tools and prompts to the model.[2]

Chen viewed terminal access as extending coding agents beyond software development. He said he used Codex to organize photos from his desktop into a folder, and offered analysis of large collections of CSV files as another application. His criterion was whether a task could be expressed through files and command-line tools.[2]

Earlier in 2025, Chen addressed a more specific integration problem: preserving a reasoning model’s state while it calls an external tool.[3]

Chen authored Better performance from reasoning models using the Responses API, dated May 11, 2025. Using o3 and o4-mini, he explained why a function-call round trip differs from an ordinary subsequent conversational turn. When the model requests a tool, the application should preserve its reasoning items through previous_response_id or explicitly return them in the next input alongside the tool result. His weather example retained the complete response.output before adding the function’s answer.[3]

The tutorial reported about a 3% SWE-bench improvement from including reasoning items with the same prompt and setup, without specifying whether that meant relative improvement or percentage points. It also reported cache utilization rising from 40% to 80% in OpenAI’s tests when switching from the Completions API to the Responses API. These were reported test results, rather than guarantees for other applications; the recipe is now marked archived.[3]

For stateless workflows, including Zero Data Retention use cases, Chen demonstrated requesting reasoning.encrypted_content and passing the encrypted reasoning items back in later requests. This preserved usable reasoning state on the client side without exposing raw chain-of-thought tokens; the tutorial separately demonstrated user-facing reasoning summaries.[3]

Before his documented OpenAI work, Chen had built an AI application for insurance and legal workflows through RiskAngle.[4][5]

Chen cofounded RiskAngle with William Hu. Founded in 2024 and part of Y Combinator’s Summer 2024 batch, the company applied AI to summarizing unstructured medical records for insurance companies and law firms.[5][4]

In a reflection published November 8, 2024, Chen said both his parents were life-insurance brokers. Navigating insurance claims following a serious skiing accident had shown him the overhead involved in medical-record review, reinforcing his belief in RiskAngle’s work. He urged founders to prioritize customers over investor pitches and described rejection as useful feedback: “Each ‘no’ forced us to rethink and refine.”[6]

Chen was born in Shanghai, raised primarily in Vancouver, Canada, and earned a bachelor’s degree in computer science from Columbia University.[5]