The posting, in Anthropic's own words
archived Oct 20, 2026About Anthropic
Anthropic’s mission is to create reliable, interpretable, and steerable AI systems. We want AI to be safe and beneficial for our users and for society as a whole. Our team is a quickly growing group of committed researchers, engineers, policy experts, and business leaders working together to build beneficial AI systems.
About the role
Anthropic is looking for a Staff+ Software Engineer to work on the vector search infrastructure behind Claude. Retrieval is how Claude learns about the world. Web search across claude.ai, the API and agentic surfaces, memory across conversations, and the reinforcement learning workloads that train our next models all depend on fast, cheap, high-recall nearest-neighbor search over billions of vectors. You'll join a small team responsible for the vector database layer that serves Claude's retrieval workloads: its architecture, performance, scalability and cost. This is a foundational role. The team is small, core architectural decisions are still open, and the engineers who join now will shape how the system evolves. The role sits in our Databases organization and works day to day with the Search Platform team, which owns web search indexing, retrieval and serving, and the Memory & Search team, which owns search and memory inside claude.ai. We're looking for someone who has built vector or search systems that ran in production, at scale, with real users and a real on-call. You'll be hands-on in the code, and you'll also be the person deciding how the pieces fit together.
Read the full posting ↓
Key responsibilities
Own the core of Anthropic's layer: index structures, storage layout, query execution and the serving layer Own production readiness - benchmarking against exact search, capacity planning, onboarding new workloads, and the operational tooling to run it at scale Drive the recall, latency and cost trade-offs - quantization, filtered search, index build and refresh strategy, and how compute and storage are separated and scaled Scale throughput by an order of magnitude as research and training demand grows, and work out whether the bottleneck is architecture, hardware or engineering bandwidth Extend the retrieval stack toward hybrid retrieval - lexical search and fused ranking alongside Set the evaluation methodology for the system - recall targets, latency objectives, cost per query, and the regression detection that keeps them honest Set technical direction for a growing team - lead design reviews, make architectural decisions others build on, and mentor the engineers joining the effort Share the on-call for the systems you own