Barr Yaron is a partner at Amplify Partners, investing in technical founders building AI, data infrastructure, and developer tools. A former data scientist and product manager, she brings an operator’s perspective to the State of AI Engineering, her annual research into how practitioners choose models, control agents, evaluate outputs, and manage the economics of production AI.
Yaron studied mathematics at Harvard and earned an MBA from Stanford Graduate School of Business. She worked in data science at Facebook and eBay and contributed to digital strategy for Beyoncé’s team after proposing that it hire a data scientist.
While working at Facebook in Israel, she founded Women of Startup Nation, interviewing more than 150 women in technology and helping organizations improve their recruiting pipelines. She subsequently developed an accelerator for women founders in partnership with Google and helped launch Stanford’s 20|20 Fund, a student investment community for aspiring entrepreneurs.
Before entering venture capital, Yaron joined dbt Labs as a product manager working on metadata, platform infrastructure, APIs, and webhooks. Amplify hired her as a principal in 2022; she is now a partner, with interests spanning companies including Axiom Bio, Chai Discovery, David AI, Inception, and Gradium. Her Barrchives interview series explores how founders and operators at companies including Datadog, Temporal, and Factory build technical products and teams.
What her research reveals about production AI
AI engineering is a discipline, not a job title. Yaron’s 2025 practitioner survey included 500 respondents; her 2026 survey, developed with Notion and Vercel, reached 1,048. Its participants included engineers, founders, product leaders, and researchers; many experienced software developers had only recently begun working with AI.
Multi-model production infrastructure beats model tribalism. In 2026, 87 percent of surveyed teams used multiple models, and more than 90 percent of open-weight-model users also used closed models. Teams selected models for quality, tool use, and cost while consolidating their surrounding infrastructure. Three-quarters reported that costs influenced their AI ambitions, making token consumption a product and monitoring concern.
Agent permissions are advancing faster than agent controls. Among respondents using agents, the proportion reporting write-enabled systems rose from 52 percent in 2025 to 89 percent in 2026. The primary safeguards remained human-in-the-loop approvals and permission gating, while persistent context, hallucinations, and evaluation remained unresolved operational challenges.
Evaluation must reflect real work. Yaron’s analysis of AI product development emphasizes domain expertise, actual user feedback, and product-specific evaluation over generic benchmarks. Her surveys also trace the tradeoffs of cheaper experimentation: broader participation in building software alongside heavier review burdens and weaker codebase comprehension.
Who is the most influential AI researcher? In Frontier Feud, a plausible answer scores only if it matches a survey of 100 AI engineers. Barr Yaron tests the builders’ knowledge of their peers through trivia, model choices, and Fast Money.
A survey of 500 respondents shows how AI engineering spans job titles, use cases, and customization methods—and where evaluation and production practices still lag.
AI engineering teams are combining more models, granting agents write access and shipping more experiments, while cost, evaluation and maintenance increasingly shape what they can build.
