Ilan Bigio builds research tools at OpenAI and helped create Swarm, the experimental multi-agent framework whose ideas informed the OpenAI Agents SDK. His work turns advanced model capabilities into practical developer tools built around explicit actions, observable state, and lightweight orchestration.
From early GPT experiments to OpenAI
A Brown University graduate, Bigio spent three years engineering on Google’s YouTube team before joining OpenAI in 2023 as a solutions architect. He had begun experimenting with GPT-3 during its 2020 beta and created Shell AI, an open-source terminal assistant he subsequently rewrote in Go.
At OpenAI, he helped customers translate model capabilities into working applications before moving into developer experience. For one latency-sensitive assistant with approximately 120 functions, he generated synthetic training examples from function schemas, used GPT-4 to label customer inputs, and fine-tuned GPT-3.5 to improve tool selection and argument generation. He also built the phone-ordering demonstration shown at OpenAI’s 2024 DevDay and contributed to Codex-related projects. His personal website identifies his current focus as building tools for research at OpenAI.
Function calling as the foundation of practical agents. Bigio treats an agent as a loop: the model requests an action, application code executes it, and the result returns to the conversation. Retrieval, memory, application state, and delegation can all emerge from that structure. His hands-on agent workshop demonstrates dynamically generated tools while warning that executing model-written code introduces serious security risks.
Routines, handoffs, and lightweight agent orchestration. In an OpenAI guide to orchestrating agents, Bigio describes workflows defined by instructions and tools, with explicit handoffs between specialists. He was a core contributor to Swarm, whose repository identifies the OpenAI Agents SDK as its production-ready successor. His advice is to introduce multiple agents only when evaluations, tool volume, or other concrete requirements justify the added complexity.
Asynchronous delegation without freezing the conversation. Bigio separates responsive interaction from slower background work: an agent creates a task, returns its identifier, continues helping the user, and later retrieves the result. This allows faster conversational models to delegate difficult reasoning without blocking the interface.
Evaluation-first model optimization. Bigio distinguishes supervised fine-tuning for constrained imitation and classification, direct preference optimization for comparative qualities such as tone, and reinforcement fine-tuning for difficult tasks with reliable graders. His model-optimization workshop emphasizes that retrieval solves missing-information problems better than fine-tuning, while an unsuccessful email-classification experiment illustrates how subjective labels and inconsistent grading undermine reinforcement learning. His practical sequence is straightforward: start with a capable model, improve the prompt, measure performance, and fine-tune only when clean data and evaluations warrant it.
Fine-tuning can make a small model a better specialist, but the training signal determines what it learns—from function calls and banking labels to preferences, chemistry, and failed email triage.
