Goldman Sachs Is Pulling Claude Code from Its Bankers and Traders

Goldman Sachs has limited Anthropic’s Claude Code and Microsoft GitHub coding tools to engineers, ending most bankers’ and traders’ access under guidance updated in October. Traders had used the agents to monitor client activity, model markets and positions, and create written and visual materials. The restriction follows a broader review of productivity, cost and risk, though “strats” - staff spanning trading, quantitative research and software engineering - retain access. Goldman’s own AI assistant, which draws on multiple models, remains available firmwide. The move reverses a more permissive approach praised in July by CIO Marco Argenti, who had described nontechnical employees building prototypes for engineers to refine.

The central concern is that coding agents can inspect files, generate and run programs with the user’s permissions, while employees without software training may deploy unreviewed code involving client data or trading decisions. Such tools can bypass testing and approval processes, mishandle unusual cases, and undermine information barriers that prevent sensitive deal knowledge from reaching traders. Goldman had separately restricted Claude for Hong Kong bankers in April over contract terms, while continuing to work with Anthropic on engineering-run back-office agents. The timing is notable: Goldman is a lead underwriter for Anthropic’s planned IPO. The account supports restricting production coding to people able to review it, but criticizes Goldman for withdrawing access without a clear replacement or systematically collecting traders’ useful applications. It argues that firms should assign engineers responsibility for agent-built tools and apply existing data, compliance and software controls to AI-generated code.