Goldman Sachs Restricts AI Coding Tools to Software Engineers

Goldman Sachs Group Inc. has revoked access to third-party artificial intelligence coding tools for most of its traders and investment bankers, restricting platforms such as Anthropic PBC's Claude Code and Microsoft Corp.'s GitHub strictly to its engineering workforce.

The revised policy, updated in October guidance and detailed in reporting by Bloomberg News and analysis from thestateofai.com (https://www.thestateofai.com/news/goldman-sachs-claude-code-traders-bankers), marks a sharp retreat from an open experimentation phase across the Wall Street firm. Under the updated mandate, autonomous coding agents are limited to Goldman's roughly 12,000 developers out of roughly 47,000 global staff. Quantitative strategists known as "strats"—specialists who bridge trading desks, quantitative analytics, and software development—also maintain access.

Frontline personnel had grown accustomed to deploying coding tools to monitor customer flow, assemble quantitative market models, and generate research collateral. The abrupt restriction frustrated desk staff who had automated bespoke workflows and now face traditional engineering request queues to build or update utilities.

Goldman spokesperson Nick Carcaterra defended the narrower deployment as an intended step in the firm's governance cycle, explaining that the bank initially introduces selected systems across departments to identify viable use cases. Following an assessment of productivity and risk management, Carcaterra said, the tools are assigned to staff equipped to handle them over the long term.

The retrenchment contrasts with comments made in July by Goldman Chief Information Officer Marco Argenti, who publicly celebrated non-technical employees engaging in "vibe coding." At the time, Argenti described how commercial teams built rough software prototypes with AI agents before handing them off to professional developers for testing and refinement.

Permitting non-developers to execute self-generated code introduced severe operational and compliance hazards into trading environments. Unlike traditional software releases, scripts generated by end-user prompts frequently run on local terminals without code review, version control, or automated security testing. For an investment bank governed by stringent regulatory information barriers between private dealmakers and public trading operations, unmonitored scripts processing live order books or client data present significant compliance liabilities.

The move follows a geographic restriction imposed in April, when Goldman pulled Anthropic's Claude from its internal platform in Hong Kong after determining the model was not officially supported in the territory under its commercial contract. Other models, including OpenAI's ChatGPT and Alphabet Inc.'s Gemini, remained active for staff in that office.

The internal restrictions arrive at an awkward commercial moment: Goldman Sachs and Morgan Stanley are serving as lead underwriters for Anthropic's initial public offering, according to an S-1 registration statement filed on October 26. While commercial bankers prepare to market shares in the AI research lab ahead of a planned November listing, most of them will no longer be permitted to run its flagship coding agent on their desks.