Goldman Sachs pulls Claude Code from traders after letting them build AI tools
Goldman Sachs has restricted Anthropic's Claude Code and Microsoft's GitHub to its engineers, cutting off most traders and bankers who had used the tools to track client flows and build market models, according to Bloomberg News as reported by thestateofai.com.
At Goldman, traders had been using it to monitor client trading flows, model market moves and positions, and assemble written and graphic material. The bank employs about 12,000 engineers among roughly 47,000 staff. Argenti told Bloomberg in July that traders and bankers built tools with AI and passed them to engineers to be improved, letting people closest to a problem build a first version without waiting for engineering time. That arrangement also meant thousands of employees without engineering training were running software that writes and executes code on the same machines they use to handle client information.
The October guidance limits Claude Code and GitHub to engineers. Strats, who work across trading, quantitative research and software engineering, still have access, according to a person familiar with the matter. The change frustrated some traders, another person familiar with the matter told Bloomberg.
Goldman had already removed Claude from its internal AI platform for Hong Kong bankers in April. The Financial Times reported that the bank read its contract with Anthropic strictly, and an Anthropic spokesperson said Claude had never been officially supported there. ChatGPT and Gemini remained available. JPMorgan removed Claude from its Hong Kong approved list in June. Those earlier restrictions were geographic and contractual; the October restriction is based on role.
Goldman's in-house AI assistant, which uses models from several companies, remains available across the firm. The restriction applies to third-party agents that can write and run code by themselves. The bank also continues to use Claude-based agents for trade accounting and client onboarding that were built with Anthropic engineers; those agents sit in the back office and are run by Goldman's engineering teams. Goldman spokesperson Nick Carcaterra said the bank makes certain AI tools available to staff "to determine the best use cases" across the firm.
The system at issue is a bank's control architecture for software that touches money and client data. Banks build information barriers between groups—between bankers working on deals and traders who could trade ahead of them, for instance. Goldman constructed its internal AI platform around these controls. A tool an engineer builds moves through review, testing, and deployment gates before it reaches production. A script a trader assembles with Claude Code does not.
Claude Code reads files, writes programs, and runs them with the user's permissions. On a trading desk, that user has live market access and can book client trades. If the agent generates a bad order, the trader's account executes it. The review gap is structural: engineering workflows assume peer review and staging environments; trader workflows assume the human in the loop is a licensed professional making intentional decisions, not delegating to an opaque agent. The bank's platform was not designed to inspect and sandbox code that traders write on the fly.
The design decision that enabled this outcome was the choice to let non-engineers run code-generating tools on production workstations without inserting the engineering controls between prompt and execution. The trade-off was speed versus safety: domain experts could build without engineering bottlenecks, but the bank accepted that their output would bypass the review chain. When management reassessed risk and productivity, they retracted the privilege rather than extend the controls. The cost of the reversion falls on traders, who lose a tool they had begun to rely on, and on the bank's public positioning, since its bankers will market Anthropic shares while forbidden from using its flagship coding product.
A commenter on the Hacker News discussion noted the likely mechanism: Goldman "is very unlikely to have decided proactively to cut off traders access," and instead "there was an operational loss." The reasoning runs that only traders send orders and book trades, Claude Code would probably have been running as the trader under their account with minimal review, and eventually an agent entered a bad order or trade. Whether or not a specific loss triggered the policy, the architecture made such an event predictable. If a system grants autonomous coding agents the same market permissions as human traders, and strips out the engineering review layer, then erroneous or hallucinated code can move directly to financial execution.
One commenter suggested designing programming environments "so they cannot hurt themselves" and "great abstractions that the bankers' agents can use." That would mean engineering teams building guard-railed interfaces or sandboxed APIs, not merely handing traders Claude Code. The open question is whether Goldman will invest in that intermediary layer or simply maintain the division between those who code and those who request code from engineers.