LangChain interviews Benchling's Head of AI on building scientific agents
Nick Larus-Stone, Head of AI at Benchling, discusses how the R&D data platform built Benchling AI — an agent-backed intelligence layer for scientists — and where the coding-agent playbook holds up or breaks down in scientific contexts.
Score breakdown
Study Benchling's approach to multi-agent design, eval without clean benchmarks, and cross-model answer verification for a concrete blueprint on adapting agentic coding patterns to domains where outputs are hard to verify.
- 01Benchling is an R&D data platform for life science companies, founded in 2012.
- 02Benchling AI launched in October 2025 as an agent-backed chat interface for scientists.
- 03Nick Larus-Stone is Head of AI at Benchling; he joined via Benchling's acquisition of his startup Sphinx Bio.
LangChain's video features Nick Larus-Stone, Head of AI at Benchling — a life science R&D data platform founded in 2012 — discussing the design and challenges of building AI agents for scientific work. Benchling AI, launched in October 2025, is an intelligence layer with a chat interface backed by an agent that helps scientists find data, design experiments, and write reports. Larus-Stone came to Benchling through its acquisition of Sphinx Bio, the analysis startup he founded.
Larus-Stone also discusses cross-checking answers between models, the role of production traces in evaluation, and context engineering differences between SQL and file-based harnesses.
The conversation covers a wide range of technical and practical topics: how Benchling's decade-plus of structured data serves as a core advantage, the architecture underlying Benchling AI, and how multi-agent architectures are used in production. Larus-Stone also discusses cross-checking answers between models, the role of production traces in evaluation, and context engineering differences between SQL and file-based harnesses. Topics such as handling verifiable versus non-verifiable tasks, running evals without clean benchmarks, and agents that create and update their own skills are also addressed.
The discussion extends to broader questions about AI in science — where AI genuinely helps today, where it still gets stuck, why fine-tuning on biology has not beaten frontier models, and when agents might realistically discover a novel cure for disease. Larus-Stone frames understanding LLMs as closer to biology than software engineering, a perspective that shapes Benchling's overall approach to agent development.
Key facts
- 01Benchling is an R&D data platform for life science companies, founded in 2012.
- 02Benchling AI launched in October 2025 as an agent-backed chat interface for scientists.
- 03Nick Larus-Stone is Head of AI at Benchling; he joined via Benchling's acquisition of his startup Sphinx Bio.
- 04The video covers multi-agent architectures, context engineering (SQL vs. file-based harnesses), and memory via agents that create and update their own skills.
- 05Benchling cross-checks answers between models to improve output quality.
- 06Production traces are used as a key tool for evaluation when clean benchmarks are unavailable.
- 07Larus-Stone argues that understanding LLMs is closer to biology than software engineering.
Topics
Summary and scoring are generated automatically from the original article. We always link back to the publisher and never republish images or paywalled content. Last processed Jun 7, 2026 · 12:45 UTC. How this works →