At Goldman Sachs, vibe coding for traders and bankers hasn’t seemed to satisfy everyone.
In October, the group reportedly refreshed its internal guidelines to broadly ban the practice for these roles. It hasn’t formalized the move, but it reiterates its policy of experimentation: evaluate “diverse elements including productivity and risk management,” then “put the AI tools in the hands of those who are best positioned to exploit them over the long term.”
A Lesson in Enthusiasm at the RAISE Summit
While passing through Paris last July for the RAISE Summit, Goldman Sachs’ Chief Information Officer expressed enthusiasm about vibe coding.
Referring to Claude Code, Codex, Devin, and Cognition, he extolled the benefits for the business lines: the ability, with the support of strats (roles at the interface between trading, quantitative analysis, and software engineering) or “forward-deployed engineers,” to design products that aren’t production-ready, but that “look exactly like what they want.”
“We are becoming prototype-driven rather than spec-driven,” the executive summarized. A stance that, he argued, significantly reduces the IT-business back-and-forth (“a few days instead of a few weeks”).
This advocacy for vibe coding did not prevent him from admitting that “in a sense,” Goldman Sachs treats the outputs of AI models as if they “came from a junior.” He notably acknowledged the tendency of models to simulate the results of external tools if those tools were not available.
The evolution of internal guidelines does not prima facie call into question the ad hoc Claude agents that Goldman Sachs has developed for activities such as market operations accounting and client onboarding. The firm also maintains an internal “transversal” assistant that leverages LLMs from Anthropic, Google, and OpenAI.
