Goldman Sachs limits Claude Code to engineers, cutting off traders

Goldman Sachs has restricted Anthropic’s Claude Code and Microsoft’s GitHub coding tools to its engineers, cutting off most traders and bankers who had been using them, Bloomberg reported, citing internal guidance updated in October. The bank says the decision reflects “a variety of factors including productivity and risk management,” according to Goldman spokesperson Nick Carcaterra.

The move ends a period in which non-engineers at one of the world’s largest banks used AI coding agents to build their own tools. It also lands weeks before Goldman helps sell shares in Anthropic itself. What follows sets out who keeps access, what the bank has said, and why coding agents raise particular issues outside engineering teams.

Who keeps Claude Code at Goldman, and who loses it

The internal guidance limits Claude Code and GitHub to Goldman’s engineers. ETtech, relaying the Bloomberg report, describes the result as most of the bank’s employees being restricted from the tools.

There is one notable exception. Strats, the staff who work across trading, quantitative research and software engineering, still have access, according to a second person familiar with the matter cited in the reporting. A summary of The New Stack’s coverage also describes these quantitative employees as exempt.

Group	Status under October guidance
Engineers	Keep Claude Code and GitHub coding tools
Strats (quantitative staff)	Keep access
Most traders and bankers	Lose access
All staff	Keep Goldman’s in-house AI assistant

The scope is specific. 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. A coding agent here means software that does more than suggest text: it reads files, writes programs and executes them.

The size of the remaining group is large. Chief information officer Marco Argenti said at a conference in Turin last month that developers, more than 12,000 of the bank’s roughly 47,000 staff, made up the first wave of Goldman’s AI adoption, and the October guidance limits the coding agents to that group.

What traders and bankers were using the tools for

According to the reporting, traders and bankers had been using the tools to monitor client trading flows, model market moves and positions, and put together written and graphic material. The change has frustrated some traders, a person familiar with the matter told Bloomberg.

The rollback contrasts with how the bank described this work only months ago. As recently as July, Argenti was praising “vibe coding” by Goldman’s non-technical employees. In that account, a trader or someone in asset management would have an idea, build a quick prototype with AI, and hand it to an engineer to scale for production. Argenti told Bloomberg that traders and bankers had built tools with AI and passed them to the firm’s engineers, of whom there are about 12,000, to be improved.

The Decoder’s reading is that this is “no longer the plan, it seems.”

Goldman’s stated reason: productivity and risk

Goldman has not published the guidance. What is on record comes through its spokesperson. Bloomberg’s report cites the “variety of factors including productivity and risk management” given by Carcaterra, who added that the experimentation phase is over and the firm is now limiting the tools to the people “who are best suited to leverage them for the long term.”

State of AI quotes Carcaterra saying the bank makes certain AI tools available to staff “to determine the best use cases” across the firm. The New Stack summary describes the bank as weighing productivity gains against security risks when deciding which employees keep the tools.

The bank’s comments name categories rather than incidents. No source reports a specific security event at Goldman behind the decision, and Bloomberg’s reporting mentions no replacement tool for the affected staff, according to State of AI.

Why a coding agent is treated differently on a trading desk

The reported reasons are general, but coverage of the decision lays out several risks specific to finance. These are the analysis of the outlets, not statements from Goldman.

Agents run with their user’s access

Claude Code reads files, writes programs and runs them, typically with the same access as the person using it. Code an engineer writes usually passes through review, testing and deployment controls; a script a trader builds to watch client flows can end up in daily use without any of those, State of AI argues. It also notes that the vibe-coding arrangement 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.

Information barriers

Banks are required to keep information barriers between groups, for example between bankers working on deals and traders who could profit from knowing about them. Goldman built its internal AI platform around controls like these, while a tool a trader assembles with an agent was not designed with them in mind.

Supply-chain and credential risk

Replit CEO Amjad Masad, citing the Bloomberg report, says similar bans are widespread in finance and points to supply-chain attacks when asked what is driving the crackdown. A Replit security advisory by CTO Luis Héctor Chávez and engineering lead Manish Mehta cites a compromised Claude Code extension found on the Cursor marketplace as an example of exploitable developer tools. The New Stack summary also lists credential exposure, malicious instructions and compromised dependencies as risks of running agents on employee machines. Unicorn API has separately covered how Claude Code can commit secrets to public repos. None of these sources ties either issue to Goldman’s decision; they are background on the wider risk picture.

Cost

State of AI suggests cost may also have played a part, since agent usage is billed by the token, and notes that Bloomberg’s sources said productivity was part of Goldman’s review. That link is the outlet’s inference, not Goldman’s statement.

A different rule from Goldman’s Hong Kong Claude restriction

This is not Goldman’s first limit on Claude. In April, the bank removed Claude from its internal AI platform for bankers in Hong Kong. The Financial Times reported that the bank had read its contract with Anthropic strictly, and an Anthropic spokesperson said Claude had never been officially supported in Hong Kong; ChatGPT and Gemini remained available there. JPMorgan removed Claude from its Hong Kong approved list in June.

The two restrictions rest on different grounds: the Hong Kong one on location and contract terms, the new one on role.

Goldman’s wider relationship with Anthropic

The restriction does not end Goldman’s work with Anthropic. In February, coverage reported that Anthropic engineers had been embedded at the bank for six months to co-develop Claude-based agents for back-office tasks such as trade reconciliation, accounting, compliance checks and client onboarding. Those agents sit in the back office and are run by Goldman’s engineering teams. At the time, coverage reported internal test results of 30% faster client onboarding and developer productivity gains of more than 20%, figures attributed to the bank’s own tests rather than independent measurement.

Goldman and Morgan Stanley are also the lead underwriters on Anthropic’s IPO, according to the S-1 filed October 26, as reported by State of AI, which expects Anthropic to list on Nasdaq under the ticker ANT before the end of November. That means Goldman’s bankers will be marketing shares in a company whose coding agent most of them can no longer use, State of AI notes.

Goldman is not alone in tightening access. ETtech reports that Goldman Sachs, JPMorgan Chase and Uber are among major companies that have tightened employee access to AI tools. State of AI adds that JPMorgan and Schwab are blocking outside agents from customer accounts, while Barclays is moving the other way: it said in October that it expects half its developers to be using Claude Code by the end of the year, and later named developers among 5,000 job cuts.

What it means for teams building with coding agents

For developers and founders, the Goldman case is less about one model than about where agent-written code is allowed to run and who is accountable for it. Several practical points emerge from the coverage:

• Role-based access is becoming a pattern. Goldman kept agents for engineers and quantitative staff while removing them from others, and kept a separate in-house assistant for everyone. Tools sold into regulated firms may need to support that kind of split.
• Isolation is a recommended control. Replit’s advisory recommends running agents in isolated, organisation-controlled cloud environments with limited access rather than on employee devices.
• Small guardrail changes matter. Research cited in The New Stack summary found that stripping a single line from Claude Code’s instructions caused its out-of-scope actions to climb significantly.
• Ownership of agent-built tools. State of AI recommends that firms giving coding agents to non-engineers decide which engineer owns what they build before it touches production data, and that firms withdrawing the tools first find out what people were using them for.

State of AI’s view is that Goldman was right to restrict the tools but wrong in how it went about it, pulling access through a guideline update without a stated replacement. That is an opinion, and the main open question is the one it raises: the traders’ tools showed real demand, and it is not yet clear how Goldman will meet it.

What to watch next

• Any replacement for traders and bankers. Whether Goldman routes their requests through engineering or offers a more controlled agent setup has not been reported.
• Other banks’ policies. Masad says similar bans are widespread in finance; further disclosures would show whether role-based limits become standard.
• Anthropic’s IPO. The listing, expected before the end of November with Goldman as a lead underwriter, may bring more questions about enterprise use of Claude Code.
• Agent security advisories. Supply-chain issues such as the compromised extension on the Cursor marketplace are part of why firms are reviewing where agents run.