Latch

Overview

Latch develops AI agents and infrastructure for biological research. Its public research program includes benchmarks for agents working with real biological data, with projects collected through benchmarks.bio. The company also offers analysis products for research teams and scientific solution providers.

The published benchmark portfolio spans spatial biology, single-cell analysis, epigenomics, preclinical pharmacology, and biosecurity-related evaluation. Latch describes these as verifiable tasks and publishes results and research openly. This profile concerns that research and evaluation contribution to agent development, rather than treating every Latch product as training-data supply.

The scientific focus is distinctive because an agent must select and execute an appropriate analysis in context. A benchmark can test whether a system reaches a defensible conclusion from data, not merely whether it recalls terminology. Reported results should still be read with the exact task, tools, and scoring setup.

Why scientific workflow evaluation matters

Our assessment is that a successful analysis requires more than code that runs. In an illustrative data-analysis task, an agent might choose a method, apply quality checks, interpret the output, and explain what the evidence supports. A technically valid pipeline can still answer the wrong question or conceal an unsuitable assumption.

Longer workflows make those errors compound. An early decision about filtering or grouping can influence every later conclusion. Evaluations that retain intermediate artifacts can help distinguish a computational failure from a scientific reasoning failure, which is useful when deciding what training or tooling to improve.

Evaluating fit

Researchers should select benchmarks aligned with the data type and analysis they intend to automate. Inspect the input context, permitted tools, reference conclusions, and how alternative valid approaches are graded. Check whether results are reproducible under the same compute and software conditions. A strong aggregate score does not replace expert review of unfamiliar scientific questions. Teams considering commercial work should distinguish access to an analysis product from custom evaluation or research collaboration. Clarify the provenance and permitted use of supplied datasets, and reserve independent material for testing. The useful outcome is more reliable scientific analysis with an inspectable basis for each conclusion.