Goldman Limits Claude Code and GitHub to Engineers and Quants

Goldman Sachs has restricted the use of Claude Code from Anthropic and Microsoft's GitHub programming tools to software engineers and the bank's quantitative analysts, known as "strats." Traders and bankers who had been using these tools lost access, according to a Bloomberg report published on Monday (2), based on updated internal guidance from October.

In trading desks, the tools had been employed for purposes unrelated to engineering: monitoring customer trading flows and producing written materials and graphics. According to the same report, some traders expressed dissatisfaction with the change.

What the Bank Said

Spokesperson Nick Carcaterra cited "a variety of factors, including productivity and risk management." He stated that the initial access period allowed the bank to evaluate use cases across different areas, and the tools are now available to those who are "best positioned to leverage them in the long term." Goldman did not disclose productivity metrics nor specify an incident that prompted the decision.

Those who lost access have the GS AI Assistant as an alternative, an internal assistant that the bank made available to its entire workforce of over 46,000 employees in June 2025. This internal tool interacts with models from several companies, including those from OpenAI, Google, and Anthropic itself, but operates within the bank's controlled environment.

The Shift in Four Months

This decision contrasts with Goldman’s own messaging. In July, CIO Marco Argenti publicly praised what is termed "vibe coding," where employees without technical training describe in natural language the program they want, and AI writes the code. Argenti described a flow where the business area created the prototype and handed it over to engineers for refinement.

The retreat is not necessarily a contradiction. One defensible interpretation is that the bank did exactly what it claimed: it opened the tool, measured the outcomes, and concluded that the non-engineering benefits did not outweigh the risks of unreviewed code running close to client data and positions. A systemically important bank is accountable to supervisors who view any model used in business decisions as an object of governance, and a script written by a trader in an afternoon hardly passes this scrutiny.

The opposing reading is also valid. If the productivity on the trading desks were clearly enhanced by the tools, the natural response would be to implement controls, not cut access. The fact that Goldman did not publish any figures leaves both interpretations open, and anyone wishing to take from the news that programming AI "does not work" outside engineering is reading more into the evidence than it supports.

New York and Bengaluru

In the United States, this decision serves as a reference for risk committees of other Wall Street banks. The signal Goldman sends is that the question has shifted from "who can experiment" to "who is responsible for the code that the tool generates." For suppliers like Anthropic and Microsoft, this means corporate contracts measured by engineering seats, not by the entire workforce.

In India, the effect is the opposite. The Bengaluru office is described by industry media as Goldman’s second-largest engineering hub, just behind New York. If access to the most powerful tools is concentrated among engineers and quants, the task of transforming a trader’s idea into code once again falls to these teams, increasing rather than decreasing the value of offshore engineering centers. For consulting firms operating delivery centers for banks in India, this news reinforces the demand for engineers skilled in working with coding agents, rather than replacing them with business users.

For the CIO of any financial institution, the Goldman case offers a concrete roadmap: separating the general-use tool made available to everyone within a controlled perimeter from the tool that writes and executes code, restricted to those who have the obligation to review it. Goldman has not published any data from the testing phase, and this silence compels each competitor to conduct the same experiment independently, at the same cost and risk.