Goldman Sachs Retrains AI Agents in Tribal Knowledge

The bank spent roughly $6 billion on AI technology last year, reflecting a broader industry pressure to justify high-stakes investments. While peers like JPMorgan CEO Jamie Dimon frame such spending as a basic competitive requirement, Goldman is focusing on the quality of output. Developers are currently tasked with refining agents—including tools like Claude and Devin—to navigate complex internal environments, data security protocols, and specific design principles that off-the-shelf models cannot intuitively grasp.

To bridge the gap, the firm has introduced "skills," which are reusable bundles of instructions that codify internal practices. A "cloud fast track" skill, for instance, directs an agent on how to handle migrations according to Goldman’s specific internal standards. By systematizing these subjective directives, the bank aims to replace the "naive" AI model with a version that understands the firm's unique engineering tenets, such as "innovating incrementally."

This shift has fundamentally altered the daily workflow of Goldman’s engineers, who now spend less time writing code and more time auditing and refining the output of their AI counterparts. The evolution also spans human interaction; junior staff, often more adept at current AI workflows, have begun mentoring senior veterans on the technology. This peer-to-peer knowledge exchange mirrors efforts at other institutions, such as Citi’s "AI accelerator" program, as banks scramble to integrate generative tools into their core operations.