Solutions Architect for Startups at OpenAI, helping early-stage companies leverage OpenAI’s APIs and models. Serial entrepreneur with 20+ years building software products, including co-founding RescueTime (YC W08) and launching Mighty AI and MessageYes at Madrona Venture Labs.
Building for the Model Wave, Not Against It
Brian Fioca joined OpenAI in May 2024 and has become a key voice on production-grade agent architecture. Featured in the GPT-5 livestream announcement (August 2025), he advocates for “future-proof” coding agents that adapt to rapid model evolution rather than fighting it.
Current Work
At OpenAI, Brian focuses on helping startups build sustainable AI systems through:
- Harness over prompts - The agent loop, tools, and interface matter more than micro-tuning prompts for specific models
- Platform stability - Using Codex as a durable agent platform that “rides the wave instead of drowning in it”
- Production patterns - Parallel tools, context compaction, MCP support, and security sandboxing built into agent systems
His philosophy: as models improve weekly (“dozens of trillions of tokens per week”), the key is designing architectures that can swap models without rewrites.
Background
Brian is a serial entrepreneur who has founded five companies and worked with more than 10 startups. His ventures include:
- RescueTime - Co-founder (YC W08), leading time management and productivity tool
- Mighty AI - Co-created at Madrona Venture Labs, acquired by Uber (2019) for self-driving car training data
- MessageYes - Co-created at Madrona Venture Labs, acquired by Nordstrom (2018) for conversational commerce
Before OpenAI, he was a Principal AI Engineer at Pioneer Square Labs (2023-2024) and Partner/Lead Engineer at Madrona Venture Labs (2014-2016).
Philosophy on Agent Development
“The harness is the special sauce, not the prompt” - Brian emphasizes that the surface area between model, user, code, and tools matters more than prompt engineering. Custom tools can be “out of distribution” for models, and poor portability across models means focusing on flexible harness design.
Trust ceiling rises with models - Rather than building elaborate scaffolding for today’s models, design for where models are going. New models will raise the trust ceiling and make complex workarounds obsolete.
Key Resources
- GPT-5.1-Codex-Max - OpenAI’s frontier coding model with native context compaction
- Agents SDK - SDK for building agents with tool support and streaming
- Voice Agents Documentation - Real-time conversational AI technology
- Voice, Verticals & Venture Podcast - Discussion on fine-tuning, startup differentiation, and voice applications (May 2025)
About OpenAI
OpenAI develops advanced language models and AI systems, including the GPT series, Codex for coding tasks, and voice agent technology. The company provides APIs and solutions for building AI-powered applications across domains.
Conference Appearance
Event: AI Engineering Code Summit 2025 Date: November 20, 2025 Time: 11:00 AM Session: Future-Proof Coding Agents: Building Reliable Systems That Outlast Model Cycles
Brian presented with Bill Chen on building coding agents that remain reliable as models evolve rapidly. The talk emphasized harness design (prompts, agent loops, tool descriptions) over model-specific optimization, arguing that sustainable agent systems require flexible architectures adaptable to new models without complete rewrites.