Goldman Sachs Restricts Claude Code to Engineers Only

Goldman Sachs has significantly narrowed who can use Anthropic’s Claude Code within the bank, marking a turning point in how Wall Street treats autonomous coding agents in regulated environments.

The bank updated its internal guidance in October to restrict Claude Code and Microsoft’s GitHub to its engineering teams. Traders, bankers, and other non-technical staff who previously built tools with the coding agent no longer have access. The move followed a review focused on productivity and risk concerns.

The restriction signals that financial institutions are adopting a cautious approach to coding agent deployments, even as some firms expand their adoption. It also creates a complicated position for Goldman as a lead underwriter on Anthropic’s IPO, since most of the bank’s employees can no longer use the product its investment banking team will soon market to the public.

What changed

Goldman Sachs removed Claude Code and GitHub access for most staff and limited the tools to its approximately 12,000 engineers out of roughly 47,000 total employees. The change, announced through October guidance and reported by Bloomberg News, came after a review of how the tools were being used across the firm.

The bank had previously allowed traders and bankers to build software using Claude Code. Those employees used the agents to monitor client trading flows, model market positions, and generate written and graphic analysis. Strategists, who work across trading, quantitative research, and software engineering, retained their access to the tools.

This followed an earlier restriction in April, when Goldman removed Claude from its internal AI platform for bankers in Hong Kong. The bank determined that its contract with Anthropic did not permit use in that territory. Anthropic stated that Claude had never been officially supported in Hong Kong. ChatGPT and Gemini remained available to Hong Kong staff. JPMorgan took a similar step in June, removing Claude from its approved tools list in that region.

Goldman’s internal AI assistant, which uses models from multiple vendors and includes built-in governance controls, remains available across the firm. The October restriction applies specifically to third-party coding agents that can write and execute code autonomously.

Why it matters

The decision reflects real tensions between the potential of agentic AI and the operational realities of regulated financial services. Unreviewed code written by people without software engineering training creates multiple risks on a trading floor.

Claude Code operates by reading files, writing programs, and executing them with the same permissions as the user running it. When an engineer writes code with the agent, that code typically goes through formal review, testing, and deployment controls before reaching production systems. A trader building a tool to watch client flows can deploy it into daily use without any of those safeguards.

Banks must maintain information barriers between different groups, for example between deal teams and traders who could profit from advance knowledge. These barriers are built into Goldman’s internal AI platform. A tool a trader assembles with a coding agent was not designed around such controls. Engineers can only review it if the trader sends it to them, which may not happen until the tool is already in use.

There is also the question of whether users understand what the code does. A trader using Claude Code to model positions is less likely to fully grasp how the agent’s code handles edge cases, which on a trading desk can translate directly to losses. A Meta engineer acknowledged that he now releases code he cannot fully account for, despite writing software being his profession.

Cost appears to have been a factor as well. Agent usage is billed by token consumption, and companies that gave agents to all employees have seen spending rise sharply. Accenture budgets approximately 29 dollars per month per employee for Claude. Amazon shut down an internal leaderboard after employees ran agents on trivial tasks to climb it.

Goldman’s chief information officer Marco Argenti was still promoting trader-built tools in July. He told Bloomberg that non-technical employees had built tools with AI and passed them to engineers for improvement, allowing people closest to a problem to create a first version without waiting for engineering capacity. The October guidance overturned that approach with no replacement mentioned in public reporting.

What to test

Before adopting coding agents across your organization, evaluate these points:

Does your vendor provide role-based access controls and audit trails? Verify that Goldman’s experience cannot be repeated because access is enforced and logged by the tool, not just by internal policy.

How does the agent handle data that passes through it? Confirm what happens to client information, market data, and other sensitive material when it appears in prompts or generated code.

What happens when code fails or behaves unexpectedly? Test whether the agent’s output is interpretable to non-engineers and whether you can trace a failure back to its source.

Is there a formal handoff process from agent-written code to production? Do not assume informal review is sufficient. Test whether governance rules written for human engineers also apply to agent-written code.

Have you tested token consumption under realistic conditions? Get actual usage data from pilot groups before rolling out to wider populations.

Does the agent’s contract cover your jurisdiction? Goldman’s Hong Kong experience shows that territorial restrictions can apply even within a global firm.

Common questions

Does Goldman Sachs still use Claude?

Yes, but in a narrower way. Goldman built Claude-based agents for back-office functions like trade accounting and client onboarding with help from Anthropic engineers. Those agents are run by Goldman’s engineering teams and sit in the back office. Goldman’s internal AI assistant, which incorporates models from multiple vendors, remains available across the firm. The October restriction applies to third-party coding agents only.

Why did Goldman remove Claude from Hong Kong?

Goldman interpreted its contract with Anthropic strictly and concluded that the bank’s Hong Kong employees should not have access to any Anthropic products. Anthropic stated that Claude had never been officially supported in Hong Kong according to the company’s website. The restriction reflected both contract language and broader geopolitical sensitivity around U.S. AI technology and Chinese access to advanced computing tools.

Who still has access to Claude Code at Goldman?

Engineers and strategists retain access. Strategists, who work across trading, quantitative research, and software engineering, kept their permissions under the October guidance. Traders and bankers lost access.

The conclusion

Goldman Sachs restricted Claude Code not because the agent itself failed, but because autonomous code execution by untrained users in a trading environment creates unacceptable governance risk. The bank’s decision is sound on the merits. Unreviewed code from non-engineers running on systems with client data, combined with the need for strict information barriers, cannot be managed after the fact.

What was poorly handled was the withdrawal itself. In July, Goldman’s CIO was still publicly praising traders who built tools with AI. A guideline update with no replacement signals to the teams who found productive uses that they, not the tool or the process, were the problem. Their complaints point to real work the engineering team could take on. Goldman would benefit more from collecting which trader-built tools proved valuable and assigning engineers to maintain them than from letting that knowledge disappear.

Other financial institutions should learn from this. Settle who owns agent-written code before distributing agents to everyone. When someone outside engineering builds a tool with Claude Code or a similar agent, assign a named engineer to take responsibility for it before it touches production data. Existing controls around information barriers, data access, and software approval should still apply when the code came from a prompt. When a tool is taken away, record what people were using it for so the work does not vanish.

Goldman remains a lead underwriter on Anthropic’s IPO. The bank’s bankers will market a product most of them cannot use, which is awkward. It is also honest about what these tools can and cannot safely do in a regulated business.

Related reading

• Claude Code vs OpenAI Codex: Pricing and Setup
• Elephant-Goldfish plugin: spec review for coding agents
• NVIDIA Auto Ontology: Open semantic layer for AI agents