Goldman Sachs Pulls Claude Code From Its Bankers Weeks Before It Markets Anthropic's IPO

Goldman Sachs has restricted Anthropic’s Claude Code and Microsoft’s GitHub to its engineers, cutting off the traders and bankers who used the coding agents to monitor client trading flows and build market models. The internal guidance was updated in October, as State of AI reported, citing Bloomberg, after a review of productivity and risk. Strats, the hybrid trading and quant staff, keep access.

The timing is what makes this a story rather than a routine IT memo. Goldman is a lead underwriter on Anthropic’s IPO, filed in an S-1 on October 26, expected to list on Nasdaq under the ticker ANT before the end of November. For a few weeks, Goldman’s bankers will market shares in Anthropic while most of them are barred from using Anthropic’s flagship coding product. That is not a scandal. It is an unusually clean illustration of how banks are actually absorbing AI agents: as tools for the people who can supervise them, not as a productivity spray across the whole firm.

The reversal is the news

In July, Goldman chief information officer Marco Argenti was promoting “vibe coding” by non-technical staff. He told Bloomberg that traders and bankers had built tools with AI and handed them to the firm’s roughly 12,000 engineers to improve. At a conference in Turin last month, Argenti described the bank’s AI adoption in waves and said developers, out of roughly 47,000 staff, made up the first. The October guidance limits the coding agents to exactly that group.

So the same executive who spent the summer describing the upside issued the restriction. There is no replacement tool mentioned in Bloomberg’s reporting. Some traders are frustrated, a person familiar with the matter told Bloomberg. They were using Claude Code to watch client flows, model positions, and assemble written and graphic material. Goldman spokesperson Nick Carcaterra said the bank makes certain AI tools available to staff “to determine the best use cases.”

That statement is doing a lot of work. It describes a firm still experimenting, not one that has settled on a policy.

Why the trading desk was the wrong place for an agent

Claude Code reads files, writes programs, and runs them, usually with the same access as the person operating it. Code an engineer writes with it typically passes through review, testing, and deployment controls. A script a trader builds to watch client flows can enter daily use without touching any of those gates.

Then there is the information-barrier problem. Banks must keep walls between bankers working on deals and traders who could profit from knowing about them. Goldman built its internal AI platform around those controls. An agent a trader spins up was not designed with them in mind, and engineers only see the code if someone forwards it.

And there is comprehension. Last month a Meta engineer said publicly that he now ships code he cannot fully account for, and writing software is his actual job. A trader modeling positions with an agent is less likely to understand how that code behaves on edge cases. On a trading desk, an edge case is a loss.

Cost likely played a role too. Agent usage bills by the token, and firms that handed agents to everyone have watched spending climb. Amazon shut down an internal leaderboard after employees ran agents on pointless tasks to move up it. Accenture now funds $29 a month of Claude per employee. Bloomberg’s sources said productivity was part of Goldman’s review.

The Hong Kong precedent

This is not Goldman’s first Claude retreat. In April the bank pulled Claude from its internal AI platform for Hong Kong bankers, reading its Anthropic contract strictly; an Anthropic spokesperson said Claude had never been officially supported there. ChatGPT and Gemini stayed available. JPMorgan removed Claude from its Hong Kong approved list in June. That restriction was geographic and contractual. October’s is role-based, and broader.

Goldman is still working closely with Anthropic elsewhere. Earlier this year the bank said Anthropic engineers spent six months on site helping build Claude-based agents for trade accounting and client onboarding. Those run in the back office under engineering teams.

The bank is not retreating from AI. It is drawing a line between agents that write and execute code unsupervised and everything else.

Barclays is running the opposite experiment

Barclays said in October it expects half its developers on Claude Code by year’s end, then three weeks later named developers among 5,000 job cuts. Like Goldman, it is giving coding agents to engineers. Unlike Goldman, it is using them to shrink the engineering headcount. Two banks, two bets on what the agent is for: a supervised multiplier for a trained few, or a reason to need fewer of them.

The broader sector is tightening too. JPMorgan and Schwab are blocking outside agents from customer accounts this fall. The pattern across finance is not rejection of AI agents. It is a scramble to decide who is allowed to point one at production data.

What this means for AI builders

The lesson for anyone selling coding agents into regulated enterprises is that the buyer’s real question is not capability. It is custody. Who reviews the code an agent writes, who owns it when it breaks, and which controls it must clear before it touches client data. Goldman answered by narrowing the user base. That is the conservative answer, and probably the right one for a bank.

The more interesting failure is procedural. Goldman spent the summer celebrating trader-built tools, then withdrew access through a guideline update with no migration path. The traders who found genuine uses, watching client flows and modeling positions, were told implicitly that they were the risk. Those workflows still exist. Someone at Goldman still needs to do them.

Watch whether the bank builds a supervised path for non-engineers, a sandboxed agent with review built in, or simply lets the demand fade. The IPO calendar will keep Anthropic and Goldman in the same room either way.