Sarah Sachs

In a Latent Space interview with Simon Last published April 15, 2026, Sarah Sachs described managing Notion’s core AI capabilities and infrastructure team, connecting agent infrastructure and evaluation with decisions about model capability and cost. She had announced her leadership of the AI Modeling Team in July 2024. Her interview account emphasized two consequential changes: distributing agent-tool ownership across product teams and building evaluations that let those teams maintain what they ship.[2][1]

Sachs explained that Notion’s earlier reliance on few-shot examples concentrated agent development around a single shared prompt. Changes competed through example selection and ordering, and only five or six people could edit or approve that part of the codebase. Moving to goal-oriented tool definitions, supported by evaluations, allowed individual teams to own their tools. The change required the central AI team to become a platform team; she called distributing tools rather than collaborating on one system-prompt string the biggest lever in scaling their contributions to the agent.[1]

Her approach to technical priorities starts with user journeys. Sachs described Friday reviews of highly token-intensive custom-agent transcripts to understand failures and turn them into tasks. Email triage was one concrete workflow the team expected to work. PDF export illustrated another: a user’s need could justify a computer sandbox, filesystem access and code execution. She argued that concentrating on interesting tools without such a journey reduced the team’s velocity.[1]

Custom Agents also required product decisions about capability and control. Sachs identified the difficulty of explaining permissions when an agent’s Slack-channel audience differed from the group allowed to access its documents. She described a turning point when the team aligned around sophisticated operators: simplifying every interaction could obscure how the agent worked and diminish its capabilities. A later change made chat the main interface for both using and configuring an agent, with settings showing the proposed changes. Sachs said this became launch-blocking; Last recalled a delay of roughly a month. She credited three engineers from three different teams with resolving it.[1]

Sachs’s team maintains the evaluation framework, while product teams own their evaluations. In her April 2026 account, these could run through opt-in continuous integration or nightly execution, with a custom agent prompting teams to inspect major failures. This arrangement kept responsibility for shipped functionality with its builders rather than transferring every prototype to a small maintenance group.[1]

She distinguished regression tests, which must pass within an allowance for stochastic variation; launch-quality evaluations, which assess whether user journeys meet release thresholds; and frontier or headroom evaluations, deliberately difficult tests targeting about a 30% pass rate. That last figure described test difficulty, not deployed-agent reliability. Sachs said saturated evaluations had stopped providing useful feedback about improving models, motivating dedicated work on what the team called “Notion’s Last Exam”.[1]

Sachs advocates Model Behavior Engineering as a distinct career involving evaluation design, failure analysis and judgment about what models can do. The function grew from data specialists who worked with Last before she joined. As coding agents became more accessible, the work expanded toward building evaluation-writing agents and LLM judges. Her position is that “you don’t need an engineering background to be the best at this job”. Agents can write evaluations, but supervision still requires people who understand the intended experience and can explain failures.[1]

Sachs describes her leadership role as clarifying objectives and helping people prioritize, while leaving room for ideas demonstrated through prototypes. Repeated harness rebuilds also require, in her words, “a team that’s comfortable deleting their own code”. She credited the pre-existing openness to changing direction to Simon and Ivan, and credited designer Brian Levin with promoting “demos over memos.” For her, easier prototyping increased the need for conviction about which user journeys deserved investment.[1]

She illustrated that approach with Jimmy, an engineer on the database collections team who wanted to build image generation for Notion cover photos and page content. Sachs described encouraging the work and providing access to Gemini, token-usage tracking, endpoints and evaluation support until it became a full project.[1]

In “Notes from Token Town: Negotiating for the Fortune 5 Million,” published March 10, 2026, Sachs argued that applied AI companies must negotiate with suppliers that also compete with them. She described Notion’s preference for provider optionality over large upfront discounts, backed by infrastructure that permits switching models. In that March 2026 essay, she reported that Auto handled about 75% of AI traffic and that model selections changed every two to three weeks as releases, tool requirements or evaluations revealed better options.[4]

Sachs distinguishes routine enterprise work from tasks that benefit from frontier reasoning. In her view, greater model capability may add little to changing a database field or triaging an inbox, while unlocking value in deep research, data analysis or incident remediation. She advocates credible open-weight alternatives for moderate workloads and reported evaluating them with Fireworks and Baseten. In the interview, she also argued for deterministic code where repeated language-model calls would create unnecessary token costs. These choices follow her emphasis on matching capability, price and latency to the task.[4][1]

Sachs argued against fine-tuning a model around rapidly changing agent tools: each tool change could require retraining and another model release, slowing development. That position coexisted with targeted fine-tuning and investment in retrieval. In April 2026, she described reworking search for agents, whose queries and information needs differed from those of humans. Inclusion within the retrieved result set and useful snippets mattered differently from click-through ranking. Query generation, ranking and retrieval were treated as one journey, rather than optimizing the embedding model in isolation.[1]

Her July 24, 2026 announcement of ZeroEntropy joining Notion extended those priorities to task-specific models. Sachs wrote that Notion had first encountered Ghita Houir Alami, cofounder and CTO Nicholas Pipitone, and their team as customers. She reported that their reranker made unified search up to 30% faster and reduced latency specifically in the reranking step by 85%, while maintaining answer quality and lowering inference cost. She announced that Alami would lead the new Model Research team and argued for AI that was fast, affordable and built for specific work.[6]

Before her Notion work, Sachs held roles in AI infrastructure and regulated financial services. Her Conffab biography places her at Tome as Director of Engineering for AI and Infrastructure, following three years at Robinhood as Head of NLP & GenAI. It describes presentation-generation and model-infrastructure work at Tome, and a transition from a BERT-powered chatbot to a compliant generative assistant at Robinhood. The biography lists earlier roles as a founding ML engineer at Sunshine, formerly Lumi Labs, and a software engineer at Google, along with an Applied Mathematics–Computer Science degree from Brown University.[5]

Sachs herself connected her years at Robinhood to bringing security into Notion projects early. In the April interview, she said late involvement caused delays and tension, while early participation helped produce a better product.[1]