Goldman Sachs pulls Claude Code from traders and bankers, engineers only

Four months ago, Goldman Sachs was one of the loudest voices for letting everyone code with AI. Now it is one of the first big banks to take that freedom away. According to internal guidance updated in October and reported by Bloomberg on November 2, 2026, the bank has limited Anthropic's Claude Code and Microsoft's GitHub coding tools to engineers. Most traders and bankers, who had used them for months, can no longer touch them.

The decision matters far beyond one investment bank. Goldman is a reference customer for enterprise AI, a firm whose technology choices are copied across Wall Street and studied by every company that wonders how far to open AI agents to non technical staff. When a bank that called itself one of the most forward leaning adopters of AI decides that coding agents belong only in the hands of engineers, it sends a signal about where the "everyone is a developer now" story runs into real world limits.

What exactly changed

The facts reported so far are short. For months, Goldman allowed bankers and traders to use Claude Code, Anthropic's agentic coding tool that runs in the terminal and can write, edit and execute code on its own, and Microsoft's GitHub coding tools for their work. The updated guidance now restricts those tools to engineers only. A group known inside the bank as "strats", employees whose jobs combine software engineering, quantitative research and trading, is exempt from the change, according to a person familiar with the decision quoted by Bloomberg.

The rollback did not land quietly. Bloomberg reports that the ban has frustrated some traders, who had used the tools to speed up parts of their job. The examples are telling: monitoring client trading flows, modeling market changes and positions, and creating written and graphic content. None of that is classic software engineering. It is the kind of everyday analytical work where a coding agent acts as a power tool for a smart non programmer, writing a quick script to slice a dataset, building a small dashboard or turning a messy spreadsheet into a chart.

Goldman's official explanation is deliberately vague. "We make certain AI tools available to our people to determine the best use cases across different parts of the firm," spokesperson Nick Carcaterra said, as quoted by Quartz. "After a period of experimentation where we assess a variety of factors including productivity and risk management, we put them in the hands of those who are best suited to leverage them for the long term."

Read carefully, that statement says two things. First, the broad access was an experiment, not a permanent policy. Second, the bank weighed productivity against risk and concluded that, for traders and bankers, the balance did not justify keeping the tools.

The July promise that did not last

What makes the reversal striking is how recently Goldman celebrated the opposite. In July 2026, Chief Information Officer Marco Argenti spoke enthusiastically about "vibe coding" by non technical employees. In his telling, a trader or someone in asset management would have an idea, build a quick prototype with AI tools and then pass it to one of the firm's 12,000 engineers to harden and scale for production.

At the RAISE Summit 2026, Argenti also put numbers on the firm's structure. Goldman has "about 45,000 employees or so", he said, with about 12,000 full time engineers and many thousands more contractors, which he summed up as "one out of three people is an engineer." He described banking as a fully digital business that is "really like constrained by the speed at which we can come up with ideas and also implement them."

That vision was attractive. If two thirds of a bank could turn ideas into working prototypes without waiting in an engineering queue, the pace of internal tooling could jump. The October guidance effectively ends that model for most of those two thirds. Prototyping now goes back to people who write software for a living, plus the hybrid strats.

Why a bank would pull back

Goldman has not published a detailed rationale, so any explanation is partly informed guesswork. But the risks of handing coding agents to non engineers inside a regulated bank are well known, and several of them fit the spokesperson's mention of "risk management".

Unreviewed code touching sensitive data. A trader who builds a script to monitor client flows is, by definition, working with client data. When the person writing that code cannot fully read or review it, the bank cannot easily tell whether the script leaks data, mishandles permissions or produces wrong numbers. Banks have spent decades building software controls, change management and audit trails. A thousand self built tools on individual laptops break that model.

Shadow IT at agent speed. Spreadsheet macros already created a famous category of risk in finance, often called end user computing. Coding agents amplify it. They can produce working code in minutes, and that code may quietly become part of how a desk makes decisions. Regulators expect banks to know which models and tools drive risk decisions. Prototypes that never reach an engineer are invisible to those controls.

Agents that act, not just answer. Claude Code is not a chatbot. It can run commands, read and write files and call tools. That is exactly what makes it useful, and it is also what makes security teams nervous. The broader industry has seen a steady stream of incidents in which coding agents or their extensions became attack paths. In an environment with client money and market sensitive data, the threshold for that kind of risk is very low.

Costs and productivity. Bloomberg notes that banks are spending millions of dollars on AI tools while trying to manage concerns about cybersecurity, costs and productivity. JPMorgan said in July it wants staff to figure out which model to use for which purpose, so the most expensive model is not used for something as simple as summarizing an analyst report. Agentic coding tools can consume large amounts of compute per task. If a trader's use case could be served by a cheaper internal assistant, the business case for a premium coding agent gets weaker.

Goldman's "productivity" wording also hints that the gains for non engineers may have been smaller than expected, or harder to measure. A quick prototype is easy to show off. Whether it turns into lasting value after an engineer rebuilds it properly is a different question.

The cognitive atrophy debate

There is a second, more cultural thread running through Goldman's AI story. In August 2026, partner Chris Churchman, who leads the bank's Marquee platform for institutional clients and co chairs its Global Banking and Markets AI working group, warned on Goldman's Exchanges podcast about the danger of outsourcing thinking.

"There's a huge danger here that in the era of AI, we outsource our reasoning to these models, and we have cognitive atrophy that stops us being able to reason from first principles ourselves," Churchman said, according to CNBC. He used junior traders as an example: they learn by handling client pricing requests under the eye of experienced risk takers. "We can absolutely automate that," he said, "but then do we get the senior traders that fully understand?"

The October guidance is not framed as a response to that warning, and there is no evidence it was. But the two ideas point in the same direction. A bank that worries about its people losing their reasoning skills has a reason to be careful about tools that let them skip the hard part of building something.

Not the first restriction this year

The coding tool decision is also not Goldman's first pullback on AI in 2026. In April, staff in Hong Kong lost access to Anthropic's Claude, including through the bank's internal AI platform, as Reuters reported citing the Financial Times. That restriction was location specific and followed a strict reading of Goldman's contract with Anthropic, which does not list Hong Kong among its officially supported markets. Other models, such as ChatGPT and Gemini, remained available there.

The two cases have different causes, one legal and geographic, the other about roles and risk. Together they show a bank that treats access to frontier AI tools as something to be granted and withdrawn, not a default right for every employee.

Goldman is still deep in AI

It would be wrong to read this as Goldman turning against AI. The bank has built its own AI assistant that combines models from OpenAI, Anthropic and Alphabet, and that internal tool is not part of the reported rollback. The change affects specific third party coding tools for specific groups of employees.

Goldman is also on the money side of the AI boom. According to Bloomberg, it has been named lead underwriter on Anthropic's initial public offering and has partnered with Nvidia to arrange major debt deals. That makes the timing a little awkward: the bank is helping to sell Anthropic to public investors in the same weeks that it restricts Anthropic's flagship coding product for most of its own staff. There is no contradiction in that, since a banker can believe in a company and still limit how a tool is used internally. But it is a useful reminder that enterprise adoption is rarely a straight line upward.

What it means for everyone else

For companies outside finance, Goldman's move is a useful data point in a debate that is far from settled. Vendors like to present coding agents as tools for everyone, from marketers to analysts. The pitch is real: non programmers can now build things that used to need a developer. But every one of those builds also creates code that someone has to own, secure and maintain.

Goldman appears to have concluded that the bank gets most of the value by putting coding agents in the hands of engineers and hybrid quant roles, and most of the risk by letting everyone else run them freely. Some companies will reach a different conclusion, especially those with fewer regulatory constraints. Others will copy Goldman's line and draw it by job function.

For individual workers, the lesson is more practical. If your job title is not "engineer", your access to the most powerful AI tools may depend less on your skills than on how your employer sorts risk. Learning to read and review the code an agent writes, not just to prompt it, may be what decides who keeps those tools when the next policy update arrives.

For AI vendors, the message is just as clear. Selling seats for coding agents to every employee in a large company will require better guardrails: controlled environments, audit logs, data access limits and easy ways to hand a prototype to an engineer. Without those, even enthusiastic customers may scale back after the experiment phase.

Open questions

Several important details are still unknown. Goldman has not said how many people lost access, how individual exceptions will be handled, or whether the restriction might be loosened again after further review. It is also unclear whether the decision was driven more by a specific incident, by cost, or by a general assessment of risk. The reports published so far include no comment from Anthropic or Microsoft on the change.

What is clear is the direction of travel at one of the most influential banks in the world: from "everyone can vibe code" in July to "engineers only" by October. The coming months will show whether other banks follow, or whether they see Goldman's caution as a chance to move faster.

Sources: Bloomberg (also syndicated via Moneycontrol), Quartz, The Decoder, RAISE Summit 2026 interview with Marco Argenti, CNBC, Goldman Sachs Exchanges, Reuters.