How Benchling Builds Agents When Even the Best AI Isn't Smart Enough

Benchling, a life sciences R&D data platform, launched Benchling AI to help scientists design experiments and find data. Instead of simply relying on a single powerful model, they use a strategy of running models from multiple model providers simultaneously.

Since each model makes different types of errors, when the results from multiple models agree, the data quality is judged to be high, and when they differ, it is considered an error, which improves accuracy.

In addition, Benchling manages agent performance in the following ways:

• Trace Review: A weekly rotating 'Fire Chief' reviews production traces and addresses issues at technical operations meetings.
• User Feedback: User 'thumbs up/down' signals are used as external signals.
• Workflow Compression: Agents reduce wait times between experimental steps, improving a day's worth of work into a week's worth of efficiency.

As a result, these AI agents are bringing major changes to the way scientific research is conducted, such as increasing the rigor of experimental design and reducing the number of experiments needed to reach a conclusion.