Goldman Sachs has pulled Anthropic's Claude Code and Microsoft's GitHub Copilot from its traders and bankers, restricting the tools to its roughly 12,000 engineers. The October guidance update follows months in which non-technical staff used AI coding agents to monitor client trading flows, model market positions, and generate presentations — all without the code review, testing, or deployment controls that govern engineer-written software. The reversal is striking because Goldman's own CIO Marco Argenti was publicly promoting the arrangement as recently as July, telling Bloomberg that traders and bankers had built tools with AI and handed them to engineers for refinement. At a Turin conference last month, he described AI adoption in waves, with developers forming the first. The October restriction concedes that the second wave — non-engineers running autonomous code agents on machines with client data access — moved faster than governance could follow. The information-barrier problem alone justifies the pullback. Banks are legally required to wall off deal-working bankers from traders who could profit on that knowledge. Goldman built its internal AI platform around such controls, but a script a trader assembles with Claude Code inherits the trader's own data access and sits outside those architectural safeguards. Agent-written code on a trading desk can reach daily production use without a single review gate. Cost and comprehension compound the risk. Agent usage is billed per token, and companies that distributed agents broadly — Amazon among them — have watched spending spike on low-value tasks. More fundamentally, a Meta engineer recently admitted he ships code he cannot fully account for, and software is his job. Traders using agents to model positions are less likely to catch how that code handles edge cases, which on a trading desk translates directly into financial loss. Goldman's relationship with Anthropic remains deep. Anthropic engineers spent six months embedded at the bank building Claude-based agents for trade accounting and client onboarding, run by Goldman's engineering teams. Goldman and Morgan Stanley are lead underwriters on Anthropic's upcoming IPO, with the S-1 filed October 26 and a Nasdaq listing under ticker ANT expected before the end of November. Goldman's bankers will spend the coming weeks marketing shares in a company whose coding agent most of them can no longer use. The competitive landscape is moving in both directions. Barclays expects half its developers to be on Claude Code by year-end and simultaneously named developers among 5,000 job cuts — using coding agents to shrink the engineering workforce rather than to gatekeep access. JPMorgan and Schwab are blocking outside agents from customer accounts. No consensus has formed on where the boundary sits, only that one is needed. The deeper lesson is sequencing. Goldman handed out powerful tools, let people find valuable uses, then withdrew access with no documented replacement path. The traders who built client-flow monitors and position models still need those capabilities. Collecting what they built and routing it through engineering review would have been generative. Issuing a guideline update and waiting for complaints to fade is not.
