Descriptions:
Kenny Workman, co-founder and CTO of LatchBio, presents the company’s work building verifiable environments and benchmarks for AI agents operating in computational biology. LatchBio started five years ago as a data tool vendor for biotech and pharma before pivoting toward agent-native infrastructure over the past two years — driven by the realization that coding agents, even without biology-specific post-training, showed surprising capability on scientific data analysis tasks.
The talk grounds the problem in scale: single-cell sequencing experiments generate 2–6 terabytes per run, spatial biology runs yield 7 terabytes, and total biological data output is growing faster than almost any other scientific domain. Standard Q&A-style benchmarks fail to capture the multi-step, interactive, data-heavy nature of real biology research, so Latch built SpatialBench — a 146-problem benchmark released in December covering spatial transcriptomics kits and analysis tasks, designed with deterministic Python graders inspired by SWE-Bench.
Workman explains the key design constraints that differentiate biological benchmarks from software engineering benchmarks: tasks must be verifiable via function, durable across valid analysis paths (since biology admits multiple correct approaches), and must require genuine data interaction rather than memorized knowledge. He notes that model capability in this domain is still low enough that breaking tasks into DAG components is necessary to get any signal at all — but early agent prototypes already demonstrate meaningful scientific reasoning, pointing toward a future where AI handles end-to-end biological research workflows.
📺 Source: AI Engineer · Published July 31, 2026
🏷️ Format: Deep Dive
