Goldman Sachs Restricts AI Coding Tools to Engineering Staff

Goldman Sachs has updated its internal guidance to restrict the use of AI coding agents, such as Anthropic's Claude Code and Microsoft's GitHub, exclusively to its engineering staff. This move effectively bars traders and bankers from using these tools to build market models or monitor client flows, a shift that follows a period of experimentation where non-technical employees were encouraged to build their own software prototypes.

This image provides a visual context for the corporate environment of Goldman Sachs as it implements new internal AI policies.

From 'Vibe Coding' to Restricted Access

As recently as July, Goldman's Chief Information Officer Marco Argenti was publicly praising 'vibe coding'—a process where non-technical employees use AI to build first-version tools and pass them to engineers for refinement. This approach allowed those closest to specific business problems to bypass engineering backlogs and create functional prototypes quickly.

The October guidance update marks a pivot from this open experimentation. While 'strats'—employees working at the intersection of trading, quantitative research, and software engineering—retain access, the majority of the bank's 47,000 staff members are now barred from using these specific coding agents.

“The bank barred its Hong Kong bankers from Claude in April and is now a lead underwriter on Anthropic's IPO.”

— Goldman Sachs Is Pulling Claude Code From Its Bankers and Traders | State of AI

Risk Management and Information Barriers

A primary driver for the restriction is the risk profile of autonomous coding agents. Unlike code written by engineers, which undergoes rigorous review and testing, a script built by a trader using an agent might enter daily use without any formal oversight. This is particularly sensitive in a regulated environment where the bank must maintain strict information barriers between deal-making bankers and traders.

This visual supports the discussion on the risks and operational trade-offs associated with autonomous coding agents.

Furthermore, there is a concern regarding technical literacy. A trader using an agent to model positions may not fully understand how the generated code handles unusual market cases. In a high-stakes trading environment, such lack of transparency in the underlying logic could lead to significant financial losses.

Operational Trade-offs and Partnerships

The move creates a friction point for staff who had already integrated these tools into their workflows. While the bank continues to work with Anthropic to build back-office agents for trade accounting and client onboarding, those specific tools are managed by engineering teams rather than being deployed directly to the trading floor.

The bank's decision highlights a broader industry trend where firms are balancing the desire for AI-driven productivity against the costs of token usage and the complexities of data entitlements. While Goldman is pulling back on these specific tools, it remains a major financial partner for the very companies providing the technology.