Goldman Took Claude Code Back From Its Traders. Who Owns the Code They Wrote?
Goldman Took Claude Code Back From Its Traders. Who Owns the Code They Wrote?
Goldman Sachs has limited Anthropic's Claude Code and GitHub to its engineers, according to internal guidance updated in October and seen by Bloomberg News. Bloomberg's report, published on 2 November, says the change followed a review of productivity and risk concerns. Some traders were frustrated. They had been using the tools to monitor client trading flows, model market changes and positions, and put together written and graphic material.
The detail that matters is not the reversal. It is what the reversal leaves behind. For months, people whose job was trading or client coverage had software they built themselves with a coding agent. The new rule says who may use the tool. It does not answer who was responsible for what the tool produced.
Every company handing a coding agent to staff outside engineering will meet the same question. Goldman met it at scale, in public, with a regulator looking at the same floor.
The arrangement was useful, which is why it spread
In July, Goldman chief information officer Marco Argenti described "vibe coding" by non-technical employees as a success. Traders and bankers built tools with AI, then passed them to one of the firm's 12,000 engineers to improve, he told Bloomberg. The appeal is obvious: the person closest to a problem ships a first version without waiting in an engineering queue.
At a conference in Turin on 9 October, Argenti put more than 12,000 of the bank's roughly 47,000 staff in the first wave of AI adoption. He described developers as becoming managers of managers, delegating work to agents that can create their own sub-agents. He also said the bank assumes its models will make mistakes and builds the environment to limit the damage, comparing it to removing the sharp edges in a kindergarten rather than handing every child a safety policy.
That is a coherent operating model, and it is entirely about the people who write and review code for a living. The traders were outside it. Bloomberg does not say how many of their tools ever completed the handoff to an engineer. The gap between "built it" and "someone signed for it" is where the trouble sits.
A coding agent arrives with the desk's permissions
Claude Code reads files, writes programs and runs them, usually with the same access as the person at the keyboard. As State of AI's analysis puts it, code an engineer writes with an agent normally passes through review, testing and deployment controls before it reaches anything important. A script a trader assembles to watch client flows can end up in daily use without passing through any of them.
Banks are also 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 with controls like those in mind. A tool a trader put together with an agent was not designed around them, and an engineer can only check it if somebody sends it over.
None of this makes the traders reckless. It makes them normal. A person with a real problem and a capable tool will build the shortest path to a fix, and the shortest path rarely includes a change advisory board. The same thing happens in a sales team tracking renewals in a spreadsheet, a support lead scraping ticket exports, or a finance analyst stitching together a reconciliation script.
A role ban is a fast control, not a durable one
The October guidance limits the tools to engineers, with "strats" (the staff who sit between trading, quantitative research and software engineering) still allowed. Goldman's internal AI assistant, which combines models from OpenAI, Anthropic and Alphabet, remains available across the firm. In State of AI's reading, the restriction targets third-party agents that can write and run code by themselves.
It is also the second restriction of its kind this year. Earlier, Goldman limited Claude for staff in Hong Kong, a location-specific rule. The new one is based on role. Both are legitimate controls. Neither is the same as owning the output.
Revoking a license does not unbuild a tool. The trader who built a flow monitor still needs to monitor flows, and will now either wait for engineering or rebuild the work in the approved assistant. Bloomberg notes that Goldman is a lead underwriter on Anthropic's IPO, which makes the restriction a decision about roles and controls rather than about the vendor. The bank still wants the models. It no longer wants them in the hands of people it has not trained to review what they generate.
What ownership looks like
State of AI's editorial take is blunt and correct: settle who owns agent-built code before handing agents out, and have a named engineer take responsibility before it touches production data. That is one sentence, and it implies a short list of things to put in place.
- Name a person, not a team. "Engineering" cannot answer for a script. A specific engineer should accept the tool before it uses live data, and that acceptance should be recorded.
- Keep the review path. Agent-written code goes through the same review, tests and approval as hand-written code, including the 40-line script that someone swears is only for personal use.
- Register the tool. Keep a list of what each agent-built tool reads, what it writes, who runs it, and which system it touches. A list a risk officer can check beats a promise that nobody would build anything risky.
- Keep entitlements on the data side. The tool should inherit only the access the job needs, not everything the builder can see. Scope is a data decision, not a prompt.
- Plan the retirement before the rollout. Record what people built and why. When access is withdrawn, the use cases should survive as a backlog item, not disappear with the license.
The common thread is that these are ordinary software controls, applied to software that arrived by prompt. Information barriers, data entitlements and software approval rules do not stop applying because the code was generated in an afternoon.
The question before the next license
Goldman's traders were watching client flows and modeling positions. They will still need to do both, and the bank will get more from collecting those use cases than from waiting for the frustration to fade. A blanket restriction is a defensible first move in a regulated business. It is a poor final one.
Before the next coding agent license goes out, ask who owns the code it will produce. If the answer is a person with a review queue and a name on the record, the rollout can scale. If the answer is "whoever built it", the restriction is already being scheduled, and it will arrive as a surprise.
If you are working out how to give business teams agent capability without losing track of what they build, talk to us.
Sources
- Goldman Restricts Traders And Bankers From Using Some AI Tools (Bloomberg via Financial Advisor Magazine, Todd Gillespie, 2 November 2026)
- Goldman Sachs Is Pulling Claude Code From Its Bankers and Traders (State of AI, 2 November 2026)
- Goldman's Marco Argenti says AI turns developers into managers of managers (The Next Web, Ana Maria Constantin, 9 October 2026)
- Hacker News discussion (8 November 2026)
