Nikunj Handa

Nikunj Handa leads product for the OpenAI API team, working with research and post-training teams to bring new model capabilities to developers. Before OpenAI, he worked at Stripe. His interests include the building blocks developers need to create agents that run for longer and interact with tools.[1][2]

Handa’s team works on capabilities such as asynchronous tool calls, which let a model continue reasoning while a tool runs, and mid-turn steering, which allows a developer to inject a message during a run. In the Latent Space interview, he describes exposing capabilities after they are trained into the harness and ready for developers to use.[1]

He has focused on reducing latency and making long-running conversations cheaper through prompt caching. Handa describes pre-warming a prompt cache before an expected request and using cache diagnostics to find where reuse drops off. For conversations that grow too large, he explains two approaches to compaction: a server-side threshold or an explicit call controlled by the developer.[1]

Handa credited Jev with inspiring OpenAI’s work on the Decisions API, which aims to make classification and other decisions fast enough to sit inside interactive systems. He described an internal prototype from the inference and infrastructure teams and said the team was working toward its latency target. Among the early uses he discussed was classifying support tickets.[1]

At Stripe, Handa saw products built on top of core payment primitives. He sees a similar design question for AI: how much should a platform provide, and how much control should remain with the builder? He points to the agent harness and possible higher-level storage or memory objects as examples of that tradeoff.[1]

what’s the right amount of flexibility to give in that?[1]

Handa also works on the developer experience around APIs and SDKs. He has highlighted the importance of SDK design and its feedback loop with the underlying API.[2][3]