Dominik Kundel

Dominik Kundel works on Developer Experience & SDKs at OpenAI. He designed and implemented the OpenAI Agents SDK for TypeScript with Kazuhiro Sera, wrote practical guidance for developers integrating gpt-oss, and has described his own use of Codex to rebuild demonstrations and handle concurrent engineering tasks. His personal website identifies him with the GitHub and Twitter handle dkundel.[2][7][4][6]

Kundel and Sera’s TypeScript Agents SDK work brought TypeScript support for AI agent development alongside capabilities for realtime voice agents. In a June 3, 2025 announcement, Sera described weeks of intensive collaboration and praised Kundel’s leadership and software architecture. Latent Space’s AI Engineer World’s Fair recap identified Kundel with the SDK launch for his workshop.[3][1]

In his one-year OpenAI reflection, published in February 2026, Kundel identified powerful voice agents using the Realtime API as a focus of their joint design and implementation. He also said he was working on the Codex app for Windows. Kundel joined OpenAI around February 2025.[7]

Kundel’s August 5, 2025 OpenAI Harmony Response Format guide explains the conversation format used to train gpt-oss: message roles establish an instruction hierarchy, while assistant channels separate user-facing answers, reasoning, and tool-related messages. For developers building their own inference systems, the guide distinguishes an ordinary message ending from signals that finish a response or hand control to a tool. It recommends the Harmony renderer to format messages and convert them into tokens; developers using a provider that handles formatting do not need to implement this themselves.[4]

Separately, Kundel reported working with researchers on the gpt-oss release and building a Rust library with the Harmony renderer.[7][4]

For his Open Models talk at DevDay 2025, Kundel decided to replace a Streamlit fine-tuning demonstration because it was confusing for the audience and had difficult behavioral bugs. In his October 10 account, he described taking screenshots, creating an initial design with v0, and downloading a mock Next.js app. He then asked the Codex IDE extension to build a FastAPI server that preserved the Streamlit app’s functionality and connected it to the new frontend. He recalled returning from lunch to a working application and subsequently adding pages to explain the demonstration better: “Without Codex, this demo would have never landed on time.”[6]

Kundel also reported assigning Codex several independent tasks simultaneously: adding Jupyter notebook support to the gpt-oss server, refactoring and fixing his agent demo, restructuring Codex documentation, and debugging a fine-tuning run. His February 2026 reflection expressed optimism about agentic coding’s effect on software development: “It's never been a better time to be a developer! You can just build things.”[6][7]

Before OpenAI, Kundel led Product & Design for Twilio’s Emerging Tech & Innovation organization, where his team worked on customer-aware AI agents. This earlier work provides context for his public guidance on the security boundaries between models and the APIs they can call.[2][5]

In “Rogue Agents: Stop AI From Misusing Your APIs”, dated October 10, 2024, Kundel argues: “We need to treat LLMs like untrusted clients, similar to how we approach browsers and mobile applications.” His shopping example makes the boundary concrete: a backend should receive an item identifier and look up its price, rather than accept a price supplied by the client. He recommends runtime validation, authentication, authorization, rate limits, least privilege, data minimization, and threat modeling for model-connected APIs.[5]

Kundel describes his team’s Twilio AI Assistants safeguards through an SMS example. TypeScript tool schemas validate model-generated inputs at runtime. Signature, session, and identity headers are generated outside the model’s flow, allowing a proxy API to authenticate requests, enforce authorization and rate limits, and validate data before calling the SMS API. The trusted identity can also determine a recipient’s phone number through a database lookup instead of leaving that choice to the model. His recommendation places enforceable controls around model actions rather than relying on prompt instructions alone.[5]