What happened
Bloomberg reports that Goldman Sachs updated internal guidance in October to limit use of Claude Code, Anthropic’s coding tool, and GitHub to engineers. The restriction follows a period in which the bank allowed bankers to use the tools for work. An exception applies to “strats,” employees who work at the intersection of trading, quantitative research, and software engineering. The source does not specify which GitHub products or configurations were covered, nor the exact date the new guidance took effect.
Some traders had used the tools to accelerate work including monitoring client trading flows, modelling market changes and positions, and creating written and graphic content. Bloomberg says the restriction frustrated some traders, citing a person familiar with the matter who was not identified because they were discussing private information. The report does not quantify how many employees had access or how much work depended on the tools.
The policy marks a contrast with comments from Goldman chief information officer Marco Argenti in July. He had described non-technical employees, including traders and bankers, using AI tools to build solutions and handing their creations to one of the firm’s 12,000 engineers for improvement. Bloomberg places the new guidance in the context of banks experimenting with AI while assessing cybersecurity risk, costs, and productivity.
Goldman spokesperson Nick Carcaterra said the firm makes certain AI tools available to determine the best use cases in different parts of the business. After a period of experimentation assessing factors including productivity and risk management, he said, Goldman puts tools in the hands of those best suited to use them over the long term. The statement does not identify the specific evidence or decision criteria behind this restriction.
The report also notes that Goldman has developed an AI assistant combining models from OpenAI, Anthropic, and Alphabet, and that the bank had earlier restricted Claude use by staff in Hong Kong. It gives no pricing, rollout plan, replacement workflow for affected staff, or indication that this latest change covers every AI product at the bank. The reported policy is specifically about the coding tools named in the article.
What the archive adds
The closest prior thread is our note on Revolut’s regulated-bank engineering model, which reported frontier tools being reserved as exceptions while internal systems served a broad engineering base: Shipping like a startup, regulated like a bank. Goldman’s restriction is a new event, not a restatement: it concerns access policy for non-engineers and does not establish that Goldman has adopted Revolut’s model.
Trend signal
This case puts pressure on the assumption that non-engineers will make up a growing share of coding-agent users. That assumption remains recorded in Coding-agent general workbench, but Goldman’s reported move shows that adoption can be narrowed by institutional risk decisions even after experimentation. One bank’s policy is not enough to establish a broad reversal.
Connections to other events
Goldman’s guidance contrasts with the earlier Revolut example in the archive, where frontier tools were described as exceptions within a system supporting about 2,000 engineers. The shared issue is not a common policy but the organizational boundary around who can use which AI tools, and how much work remains inside engineering teams.
Why it matters
For teams building with AI, the lesson is that access design is part of deployment, not an administrative detail to settle after a successful pilot. A tool may help employees prototype workflows, yet an organization can still decide that durable use belongs with a narrower group because of risk, productivity, or oversight concerns.
Goldman’s former approach, as described by its CIO, allowed non-technical staff to create solutions and then hand them to engineers for refinement. The reported change may reduce that path for many bankers and traders, while preserving access for employees whose work already combines trading, research, and software engineering. Other teams can use this distinction to plan pilots: specify who may experiment, how prototypes are reviewed, who owns production versions, and what evidence would justify broader or narrower access. The source does not tell us whether Goldman’s revised arrangement is more effective; it makes clear that experimentation alone did not guarantee broad ongoing access.
Limits and uncertainty
The report is based on internal guidance seen by Bloomberg and accounts from people familiar with the matter; it does not reproduce the full guidance. Goldman’s spokesperson confirms an experimentation-and-selection approach but does not explicitly confirm every operational detail attributed to the directive. There are no figures for productivity, risk incidents, costs, tool usage, or affected employees. The article also does not say whether restricted workers retain access to other AI assistants, whether the decision is temporary, or whether it applies beyond the named coding tools. No causal conclusion about safety or performance can be drawn from the reported change alone.
Reader takeaway
Goldman’s reported decision is a reminder that successful AI experimentation does not automatically become broad deployment. In a regulated organization, the lasting product boundary may be the combination of user role, review path, and risk owner—not simply whether a coding agent can perform a useful task. The evidence here is one company’s policy shift, not proof that restricting access improves outcomes.