Barr Yaron

Barr Yaron is a partner at Amplify Partners. She runs an annual survey of AI engineers and hosts Barrchives, interviewing technical-company CEOs and CTOs. Her public work examines how teams move AI into production: evaluating quality, deciding when to build models, and controlling agents that can change data inside software systems.[2][6][5]

Yaron authored Amplify’s 2025 AI Engineering Report, published June 24, 2025, and its 2026 report, published July 2, 2026. The latter surveyed more than 1,000 engineers building in AI in collaboration with Notion and Vercel. Justin Gage edited both reports, with Jenna Fleischer also editing the 2026 edition.[4][5]

Her surveys distinguish agent adoption from agent autonomy. The 2025 edition defined an agent as “a system where the LLM controls the core decision making or workflow.” In the 2026 report, 89% of agent-using teams said their agents could write data, compared with 52% in the previous year’s survey; most still required a human in the loop. Yaron described the leading safeguards—human approvals and permission gating—as comparatively primitive, with other control strategies scattered across respondents.[4][5]

Presenting her survey at the 2026 AI Engineer World’s Fair, Yaron emphasized the consequences of that write access: “Agents are no longer reading, summarizing, drafting,” she said. “They’re taking actions inside the systems.” Her assessment was that “Nobody has settled the control layer for agents,” according to Richard MacManus’s July 3 dispatch.[1]

The 2026 findings also informed her account of practical model selection. Forty-five percent of respondents used open-weight models, and more than 90% of that subgroup also used closed models; she interpreted open weights as a companion to closed models rather than a direct replacement. Cost regularly shaped how ambitiously 40% of respondents used AI and sometimes did so for another 36%. These findings describe the report’s fairly AI-native respondent group.[5]

Yaron also revisited an earlier expectation. In 2025, the report anticipated an audio-adoption wave because 37% of respondents not using audio intended to adopt it soon. In 2026, the share reporting that they used audio and it worked well had risen from 17% to 24%—growth she acknowledged was below what might have been expected. Meanwhile, intended adoption among nonusers rose to 56%, keeping the distinction between enthusiasm and realized use visible.[4][5]

In her February 3, 2026 essay, “How to build AI into your technical product”, Yaron synthesized lessons from her CEO and CTO interviews. She argued that domain expertise must be embedded in engineering work. Her Vanta example describes compliance experts working alongside engineers and product managers to build a golden dataset of expected inputs and desired outputs. Vanta then uses LLM-as-judge evaluations in its CI/CD pipeline to detect whether code changes improve or degrade outputs, alongside production-quality monitoring.[6]

For Yaron, evaluations also need feedback from actual use. She described Replit measuring quality through explicit user reactions and implicit signals such as rolled-back changes, and Suno learning from users’ preferences between two generated songs. Her conclusion was: “Ultimately, the best evals are living systems. They’re built from domain knowledge, refined through human judgment, and validated in the real world.”[6]

Her build-versus-buy argument centers on how product requirements and model capabilities change. In her account, Replit initially trained models because code completion demanded low latency. As its focus shifted toward agents completing whole projects, accuracy and cost became more important than shaving seconds off a multi-minute task. The company shifted to off-the-shelf models fine-tuned for particular agent actions. Yaron’s recommendation was to identify where differentiation resides—in model weights, data or surrounding software—rather than assume that owning a model is inherently advantageous.[6]

Yaron distinguishes applied AI teams using foundation models from companies that need original research because existing models cannot solve their problems. She advocates shared platform capabilities such as evaluations, inference and data pipelines for product teams, and close collaboration between researchers and engineers where research is necessary. “Research should generate product hypotheses. Products should generate research data,” she wrote, while recognizing that in deep-tech work, technical discovery may precede product definition.[6]

Before her AI surveys and podcast synthesis, Yaron worked in data science and technical product management. She studied mathematics at Harvard and applied her data-science skills at Facebook and eBay. She later worked in product at dbt Labs on platform, metadata, APIs and webhooks; Amplify’s account of her career specifically identifies metadata management as a focus. This earlier work provides context for her attention to data-driven decisions and technical products.[2][3]

While working as a Facebook data scientist, Yaron moved to Israel to spend time with extended family. She founded Women of Startup Nation, sharing the stories of technical women in Israel, and says she interviewed more than 150 women, including founders. She and collaborators subsequently launched a full-time accelerator in partnership with Google. At Stanford Graduate School of Business, she also launched 20|20 Fund with classmates to invest in and support fellow students’ entrepreneurship.[2]

Amplify introduced Yaron as a Principal on October 10, 2022. Its announcement described her choosing dbt product management after business school while continuing to meet entrepreneurs and introduce founders to Amplify. Her employer profile now lists her as Partner. The 2022 announcement also linked her historical social handle, @barnanas.[3][2]