Goldman Sachs restricts Claude Code to engineering staff

Restricting AI coding tools

Goldman Sachs has updated its internal guidance to limit the use of Anthropic's Claude Code and Microsoft's GitHub to its engineering teams. This change effectively bars traders and bankers from using these specific tools to build market models or monitor client trading flows.

The decision comes after a review of both productivity and risk concerns within the firm. While engineers remain authorized to use these coding agents, most other employees will no longer have access to software that can write and execute code independently.

The bank's spokesperson, Nick Carcaterra, stated that the firm makes certain AI tools available to staff "to determine the best use cases" across the organization. This move signals a shift toward more controlled environments for high-powered AI capabilities.

The rise of vibe coding

Before these restrictions, Goldman's chief information officer, Marco Argenti, had been publicly praising a trend of non-technical employees building their own tools. This practice, sometimes referred to as "vibe coding," allowed traders and bankers to create first versions of software without waiting for engineering time.

These employees were using AI to create scripts for monitoring positions and generating graphic materials for clients. However, this arrangement meant that thousands of staff members without formal engineering training were running software on the same machines they use to handle sensitive client information.

The bank's engineering team, which consists of about 12,000 people, was originally intended to take these employee-built tools and improve them. By restricting the tools now, the bank is moving away from this decentralized approach to software creation.

Risk and information barriers

One of the primary reasons for the restriction involves the lack of oversight for agent-written code. Unlike code written by a professional engineer, which undergoes rigorous testing and deployment controls, a script built by a trader could enter daily use without any formal review.

Financial institutions are also required to maintain strict information barriers between different groups. For example, bankers working on specific deals must be kept separate from traders who could profit from that information, and AI agents could potentially bridge those gaps.

Furthermore, there is a concern that non-engineers may not fully understand how an AI-generated script handles unusual market cases. On a trading desk, a failure to account for these edge cases could result in significant and rapid financial losses.

Contractual and cost factors

The bank's restrictions are also influenced by geographic and contractual limitations. In April, Goldman removed Claude from its internal platform for bankers in Hong Kong, a move attributed to a strict reading of the bank's contract with Anthropic.

Cost may also be a factor in the bank's decision to limit access. Because agent usage is billed by the token, many companies have seen costs climb as employees run agents on various tasks. Some firms have even shut down internal leaderboards to prevent employees from using agents for non-productive activities.

Despite these restrictions, Goldman remains a key partner for Anthropic. The bank and Morgan Stanley are currently serving as lead underwriters for Anthropic's upcoming IPO, even as Goldman's bankers market shares for a company whose coding tools they are no longer allowed to use.

The path forward for banks

Goldman's move is part of a broader trend of financial institutions tightening the reins on AI agents. Other banks, including JPMorgan and Schwab, have recently blocked outside agents from accessing customer accounts to ensure data security.

While some firms like Barclays are using coding agents to help reduce the number of engineers they employ, Goldman is focusing on role-based access. The bank's internal AI assistant, which uses models from various providers, will remain available to all staff for general tasks.

The bank's strategy suggests that while AI is useful for general productivity, the ability to write and run autonomous code requires specialized oversight. Moving forward, the firm will likely focus on using these agents in back-office roles where they can be managed by dedicated engineering teams.