Brian Fioca is an OpenAI engineer focused on frontier-model evaluation and coding agents and a co-founder of RescueTime. His career spans productivity software, startup incubation, and the infrastructure that makes advanced AI systems practical for developers.
Fioca co-founded RescueTime in 2007; the company joined Y Combinator’s Winter 2008 batch. Its software tracked how people spent time on their computers, eventually emphasizing focus, healthier work habits, and burnout prevention. During its 2021 relaunch, Fioca opposed using productivity monitoring for workplace surveillance.
At OpenAI, Fioca has worked with startups on applied AI, coding agents, and frontier-model evaluations. He is a core contributor to OpenAI’s realtime-agents repository and a credited contributor to the GPT-5 System Card.
The agent harness determines practical reliability. Fioca defines a coding agent through its interface, model, and agent harness: the layer coordinating prompts, tools, execution, permissions, and software context. Maintaining that layer becomes harder as models and APIs change.
Model habits should shape prompting. His approach to model-specific prompting accounts for behaviors acquired during training. Instructing an already-thorough model to inspect everything can unnecessarily increase latency; effective instructions work with its defaults.
Long-running agents need operational discipline. Reliable execution depends on context-window compaction, parallel tool calls, sandboxing, permissions, caching, and external-tool connections. Reusable infrastructure lets product teams concentrate on differentiated workflows.
Coding agents can build their own integrations. Fioca has described software that creates customer-specific API connectors and demonstrated a Kanban board capable of fixing its own bugs. He also helped develop frontend-design guidance for GPT-5.4, emphasizing clear constraints, visual references, and substantive content.
A coding agent depends on more than model intelligence: its tools, prompts, and runtime must fit the model’s habits without making every upgrade a rebuild.
