Until October, a Goldman Sachs trader who wanted a tool to watch client order flow, or a quick model of how a book would move if rates jumped, could build it with Claude Code. Internal guidance updated that month ended that. According to Bloomberg’s November 2 report, Anthropic’s Claude Code and Microsoft’s GitHub tools are now limited to Goldman’s engineers, and the traders and bankers who had used them for months have to ask for alternatives or fall back on the firm’s internal AI assistant.
Most of the coverage read this as Wall Street souring on “vibe coding.” That reading is too simple. The tool did what it was sold to do: it let people who are not software engineers produce working software. What a bank has trouble governing is the software they produced, and the rules a bank answers to were rewritten this spring in a way that makes that distinction sharper, not softer.
What changed, and what Goldman actually said
The facts in the reporting are narrow, so it is worth being precise about them.
- Who lost access: traders and bankers. Per the Bloomberg story as syndicated by Yahoo Finance, traders had used the tools to monitor client trading flows, to model market changes and positions, and to produce written and graphic material. Some are frustrated.
- Who kept it: engineers, and “strats,” the Goldman staff who sit between trading, quantitative research and software engineering. The directive does not apply to them.
- The stated reason: spokesperson Nick Carcaterra said the firm weighs “a variety of factors including productivity and risk management” before putting tools in the hands of people “best suited to leverage them for the long term,” as The Decoder summarized. PYMNTS quotes him adding that the firm makes “certain AI tools available to our people to determine the best use cases across different parts of the firm.”
- What it reverses: in July, CIO Marco Argenti praised non-technical staff who vibe coded prototypes and handed them to one of Goldman’s 12,000 engineers to finish, per the same reporting.
Two pieces of context make this more than a seat-count decision. Goldman is the lead underwriter on Anthropic’s IPO, so this is not a bank walking away from a vendor it distrusts; the S-1 Anthropic filed in October is partly Goldman’s work. And this is Goldman’s second Claude restriction of 2026. In April it cut Claude access for bankers in Hong Kong, a move Reuters linked to a review of its agreement with Anthropic. The April decision was about where a model may run. The October decision is about who may make things with it.
Goldman has not said which risk drove the change, and the reporting does not name one. What follows is analysis of the rulebook the bank operates under, not a claim about its internal reasoning.
The regulator scoped out the agent and left its output in scope
On April 17 the Federal Reserve replaced its fifteen-year-old model risk guidance, SR 11-7, with SR 26-2, “Revised Guidance on Model Risk Management”. Footnote 3 of the attached guidance is the sentence everyone in bank AI governance has read by now:
“Generative AI and agentic AI models are novel and rapidly evolving. As such, they are not within the scope of this guidance.”
The footnote continues that a bank’s own risk management and governance practices “should guide the determination of appropriate governance and controls for any tools, processes, or systems not covered in this document.” So Claude Code itself is not a “model” for SR 26-2 purposes. The bank decides how to govern it.
Now read the definition the guidance does apply:
“For the purposes of this guidance, the term ‘model’ refers to a complex quantitative method, system, or approach that applies statistical, economic, or financial theories to process input data into quantitative estimates.”
It excludes “simple arithmetic calculations, such as those found within spreadsheets, as well as deterministic rule-based processes and software where there are no statistical, economic, or financial theories underpinning their design or use.”
Put the two passages next to the reported uses. A script that pulls client flows into a dashboard is probably a deterministic process and sits outside the definition. A script that models how positions move when the market does is, on a plain reading, a quantitative method applying financial theory to produce estimates. The agent that wrote it is out of scope. The thing it wrote may be in.
The guidance is also explicit about which models get the hardest look. Materiality is set by exposure and purpose, and “models developed to help meet regulatory requirements or manage a banking organization’s financial risk exposures are generally considered to be of greater risk than models that are not used for such purposes.” A trader’s position model sits close to that line. And the inventory section expects banks to keep “sufficient information to understand model risks” at both the individual and aggregate level.
That is the operational problem with handing a coding agent to a trading desk. The person who decides whether an artifact is a dashboard or a model is the person who built it, in an afternoon, with no reason to ask. Argenti’s July arrangement assumed the engineer was the checkpoint: the trader prototypes, the engineer productionizes. An agent that produces something usable on day one lets the prototype go into use before the handoff happens. Each one may be harmless. The aggregate is a set of unregistered quantitative tools feeding trading decisions, which is the exact inventory gap the guidance tells banks to close.
Seen that way, the strats exemption is the most telling detail in the story. Strats are the people whose work already lives inside model governance, so their output arrives pre-registered. The line Goldman drew is not between technical and non-technical staff. It runs between people whose artifacts already enter a control process and people whose artifacts do not.
The same line, drawn outside a bank
Most companies reading this do not answer to the Fed. They have the same problem anyway, and they have had it for thirty years under another name: the finance team’s spreadsheet that quietly became the pricing engine. Coding agents make that spreadsheet faster to build and harder to see, because it now looks like software and runs on a schedule.
A role-based ban is one answer, and it is the easiest to enforce. It is also expensive, because it throws away the prototypes that were worth keeping. If you are setting this policy at your own company, three choices hold up better than a job-title list.
- Govern by what the artifact touches, not who wrote it. Anything that produces a number someone acts on (a forecast, a price, a risk figure, a customer decision) gets registered with an owner and a reviewer before it runs on live data. A formatting script does not. The test is the output’s use, which is how SR 26-2 frames materiality too.
- Separate building from running. Let non-engineers prototype against a snapshot or synthetic data in a sandbox, and require an engineering or model-risk review before anything reads production data or runs unattended. That keeps Argenti’s July hand-off model and puts the checkpoint where the agent cannot skip it.
- Enforce it in the tool, not the memo. Claude Code supports organization-deployed managed settings, and Anthropic’s documentation is direct about their precedence: “Nothing in your own settings files or
--settingsoverrides a managed key.” Permission rules, allowed models and directory limits can be pushed to every machine or account the organization controls. A policy that lives only in a PDF is a policy the agent’s user decides whether to follow.
None of this replaces the security work a coding agent needs on its own account, from vetting the mods and extensions people install to knowing that a trusted marketplace can ship a backdoored build. But those controls protect the machine. The question Goldman ran into is about the output, and it needs a record of what the agent did and what it produced, not just what it cost.
There is a fair counterargument, and Goldman’s traders are presumably making it. The case for non-developers shipping real software was always that the people closest to a problem build the best first version of the fix. Pulling the tool loses that. The better long-run answer is probably a narrower, enforced path back in for the desks that can show what they build and who reviews it.
What Goldman’s reversal settles is the order of operations. Giving a desk a coding agent is a decision about what that desk will produce. A firm that cannot say where those products are registered, who checks them, and what happens when one is wrong has not finished making the decision, however good the pilot looked.