The tool design tricks behind Benchling's AI agents | Nick Larus-Stone

<p>Nick Larus-Stone is the Head of AI at Benchling, the R&amp;D data platform that life science companies use to store and manage their experiments, samples, instruments, and analysis. Benchling has been around for since 2012. In October 2025, it launched Benchling AI, an intelligence layer with a chat interface, backed by an agent, that helps scientists find data, design experiments, and write reports. Nick came to Benchling through its acquisition of Sphinx Bio, the analysis startup he founded. In this conversation, Nick walks through what it takes to build agents for scientific work, and where the playbook from coding agents holds up and where it breaks down.</p><p>–</p><p><strong>We also discuss:</strong></p><ul><li>Why Benchling invests so heavily in getting clean data upfront</li><li>How they cross-check answers between models to get more out of each one</li><li>Why and how Benchling leans on production traces</li><li>Where AI actually helps science today, and where it still gets stuck</li><li>Why understanding LLMs is closer to biology than software engineering</li></ul><p>–</p><p><strong>Timestamps:</strong></p><p>(00:00) Intro</p><p>(01:22) What Benchling AI is, and the 14-year data platform underneath it</p><p>(04:36) Why a decade of structured data is a core advantage</p><p>(05:57) The architecture under the hood</p><p>(08:28) Similarities and differences compared to a coding harness</p><p>(11:14) Benchling’s multi-agent architectures</p><p>(14:36) Dealing with verifiable vs non-verifiable tasks</p><p>(16:19) Doing evals when clean benchmarks aren’t possible</p><p>(18:13) Context engineering: SQL vs. file-based harnesses</p><p>(22:11) Memory: agents that create and update their own skills</p><p>(25:30) What user education for scientists looks like</p><p>(30:33) Why understanding LLMs is closer to biology than software</p><p>(33:28) When will agents discover a novel cure for disease?</p><p>(44:58) The future of harnesses in science</p><p>(48:13) Why fine-tuning on biology hasn&#39;t beaten frontier models</p><p>–</p><p><strong>References:</strong></p><ul><li><a href="https://docs.claude.com/en/docs/agents-and-tools/agent-skills/overview" target="_blank" rel="ugc noopener noreferrer">Agent Skills (Claude Docs)</a></li><li><a href="https://www.benchling.com/blog/complex-questions-fast-answers-benchling-deep-research" target="_blank" rel="ugc noopener noreferrer">Benchling’s Deep Research Agent</a></li><li><a href="https://www.anthropic.com/claude" target="_blank" rel="ugc noopener noreferrer">Claude (Anthropic)</a></li><li><a href="https://en.wikipedia.org/wiki/Design_of_experiments" target="_blank" rel="ugc noopener noreferrer">Design of experiments (DOE)</a></li><li><a href="https://www.fda.gov/drugs/types-applications/investigational-new-drug-ind-application" target="_blank" rel="ugc noopener noreferrer">FDA Investigational New Drug (IND) application</a></li><li><a href="https://gemini.google.com/" target="_blank" rel="ugc noopener noreferrer">Gemini (Google)</a></li><li><a href="https://research.google/blog/accelerating-scientific-breakthroughs-with-an-ai-co-scientist/" target="_blank" rel="ugc noopener noreferrer">Google AI co-scientist</a></li><li><a href="https://www.langchain.com/langsmith" target="_blank" rel="ugc noopener noreferrer">LangSmith</a></li><li><a href="https://modelcontextprotocol.io/" target="_blank" rel="ugc noopener noreferrer">Model Context Protocol (MCP)</a></li><li><a href="https://ghuntley.com/ralph/" target="_blank" rel="ugc noopener noreferrer">The Ralph (Wiggum) Loop (Geoffrey Huntley)</a></li><li><a href="https://www.benchling.com/blog/resync-bio-and-sphinx-bio-join-benchling" target="_blank" rel="ugc noopener noreferrer">Sphinx Bio</a></li></ul><p>–</p><p><strong>Where to find Nick:</strong></p><ul><li><a href="https://www.benchling.com/" target="_blank" rel="ugc noopener noreferrer">Benchling</a></li><li><a href="https://www.linkedin.com/in/nlarusstone/" target="_blank" rel="ugc noope