Goldman Sachs has effectively revoked access to Anthropic’s Claude Code and Microsoft’s GitHub for its traders and bankers, restricting these AI coding agents strictly to its engineering department. Internal guidance updated in October, as reported by Bloomberg News, limits the tools to the bank’s approximately 12,000 engineers, ending a period where non-technical staff used the agents to monitor client flows and model market positions. This move signals a sharp pivot from the 'vibe coding' enthusiasm expressed by CIO Marco Argenti in July, replacing open experimentation with a rigid, role-based permission structure.

The Irony of Underwriting Anthropic

The timing of this restriction creates a bizarre corporate contradiction: Goldman Sachs is currently a lead underwriter on Anthropic’s upcoming IPO. According to the S-1 filed on October 26, Goldman and Morgan Stanley are marketing shares for the AI giant, with a Nasdaq listing expected under the ticker ANT before the end of November. The bankers responsible for selling Anthropic’s stock to investors are the same employees now barred from using Anthropic’s flagship coding agent, a tool that helped define the company’s recent valuation surge.

Risk Management Overrides Productivity

The decision stems from a review of productivity and risk, specifically targeting the dangers of unreviewed code on trading floors. Claude Code reads files, writes programs, and executes them with the user’s permissions, bypassing traditional software development lifecycle controls. For traders, this meant scripts monitoring client flows could enter daily use without peer review or testing, posing significant risks to data integrity and compliance. Goldman’s internal AI platform, which retains access to models like ChatGPT and Gemini for general use, was built with strict information barriers in mind—barriers that ad-hoc agent-written code often ignored.

A Precedent Set in Hong Kong

This is not Goldman’s first clash with Anthropic’s access policies. In April, the bank removed Claude from its internal AI platform for bankers in Hong Kong, citing a strict reading of the contract terms. While JPMorgan followed suit in June, Goldman’s latest move is broader, applying to all non-engineering staff globally rather than being limited by geography. The October guidance specifically targets third-party agents capable of writing and running code autonomously, distinguishing them from passive chat assistants that remain available to the wider firm.

Diverging Strategies in Finance

While Goldman restricts access, rival Barclays is expanding it, with plans for half its developers to use Claude Code by year-end, even as it cuts 5,000 jobs. Barclays is using AI agents to reduce headcount, whereas Goldman is using role restrictions to contain operational risk. Other institutions like JPMorgan and Schwab have similarly blocked outside agents from customer accounts this fall, indicating a sector-wide trend of tightening controls around autonomous AI tools in regulated environments. The divergence highlights a fundamental disagreement on whether AI coding agents are productivity multipliers or liability generators.

Key Takeaways

  • Goldman Sachs restricts Claude Code and GitHub to its approximately 12,000 engineers, barring traders and bankers from using these autonomous coding agents.
  • The policy shift contradicts CIO Marco Argenti’s July praise of 'vibe coding' by non-technical staff, signaling a pivot from experimentation to strict risk management.
  • Goldman is a lead underwriter for Anthropic’s IPO, expected to list on Nasdaq as ANT in November, creating a conflict where bankers sell shares in a company whose tools they cannot use.
  • Risk concerns over unreviewed code execution and information barriers drove the decision, while rival Barclays expands Claude Code usage to reduce developer headcount.

The Bottom Line

Goldman Sachs was right to restrict unreviewed agent code on the trading floor but wrong to withdraw access without a clear replacement strategy. Leaders must define ownership for agent-built tools and record use cases before removing them to avoid alienating productive staff.