Goldman Sachs cuts traders and bankers off from Claude Code and GitHub

Coding agents are now an engineers-only tool

Internal guidance updated in October restricts Anthropic's Claude Code and Microsoft's GitHub to Goldman Sachs engineers, Bloomberg News reported, following a review that weighed productivity and risk. Most traders and bankers at the firm lost access, though "strats" — staff who work across trading, quantitative research and software engineering — keep it, according to a second person familiar with the matter cited by Bloomberg.

The change has frustrated some traders. They had been using the tools to track how clients trade, model market moves and positions, and prepare written and graphic material. Goldman spokesperson Nick Carcaterra said the bank makes certain AI tools available to staff "to determine the best use cases" across the firm.

A reversal from the summer's vibe-coding push

As recently as July, chief information officer Marco Argenti was praising non-technical staff who built tools with AI and passed them to the bank's engineers — roughly 12,000 of about 47,000 employees — to be improved. Speaking at a conference in Turin in October, he described Goldman's AI adoption in waves, with developers as the first wave. The new guidance confines the coding agents to that group.

That arrangement let the people closest to a problem build a first version without waiting for engineering time. It also meant thousands of employees without engineering training were running software that writes and executes code on machines they also use to handle client information.

Hong Kong came first, on different grounds

In April, Goldman removed Claude from its internal AI platform for bankers in Hong Kong. The Financial Times reported the bank had read its contract with Anthropic strictly, while an Anthropic spokesperson said Claude had never been officially supported in Hong Kong; ChatGPT and Gemini remained available. JPMorgan dropped Claude from its Hong Kong approved list in June.

The April restriction turned on location and contract terms; October's turns on role. Goldman's in-house AI assistant, which draws on models from several providers, remains available across the firm. What is restricted is the class of third-party agents that can write and run code by themselves.

The Anthropic relationship deepens anyway

The clampdown has not cooled ties. Anthropic engineers spent six months at Goldman this year helping build Claude-based agents for trade accounting and client onboarding — back-office systems run by the bank's engineering teams. Goldman and Morgan Stanley are lead underwriters on Anthropic's IPO, according to the S-1 filed October 26, with the listing expected on Nasdaq under the ticker ANT before the end of November. As State of AI notes, the bankers marketing those shares will mostly do so without access to the coding agent at the centre of the deal.

Why unreviewed agent code worries a bank

A tool like Claude Code reads files, writes programs and executes them, typically with the same access as the person running it. Code an engineer writes usually passes through review, testing and deployment controls before reaching anything important; a script a trader assembles to watch client flows can enter daily use without any of that. Regulated banks must also maintain information barriers between groups, for example between deal bankers and traders who could profit from advance knowledge of those deals, and ad-hoc tools are not designed with those walls in mind. Cost may have played a part as well: agent usage is billed by the token, and Bloomberg's sources said productivity formed part of Goldman's review.

Other banks are drawing their own lines

Barclays is heading the other way on scope, saying in October that it expects half its developers to be using Claude Code by year end — and three weeks later naming developers among 5,000 job cuts. JPMorgan and Schwab have been blocking outside agents from customer accounts this autumn. Goldman's move is narrower than any of these: it limits which of its own employees may use coding agents at all.

Why it matters

Goldman has drawn one of the clearest lines yet in how heavily regulated firms treat AI: general-purpose assistants for everyone, autonomous code-writing agents only for people trained to review what they produce. The tension is instructive for any company handing out agents. Traders found genuine uses for these tools, and the bank's own leadership was celebrating that work months earlier, yet unreviewed code running next to client data is a risk few banks will carry for long. The emerging pattern looks like role-based access, clear ownership rules for agent-built code before it touches production data, and a widening gap between firms that merely restrict agents and firms like Barclays that pair them with headcount reductions.