Goldman Sachs has taken AI coding agents away from most of its bankers and traders. Internal guidance updated in October limits Anthropic’s Claude Code and Microsoft’s GitHub coding tools to the bank’s engineers, Bloomberg News reported on November 2, 2026. Strats, the staff who combine trading, quantitative research and software engineering, keep access.
The story went around again on November 8, when a State of AI write-up reached the Hacker News front page with more than 650 points. Goldman’s stated reason is short: spokesperson Nick Carcaterra cited a “variety of factors including productivity and risk management.” Below: what changed, who still has the tools, what traders were using them for, and what the decision says to any company handing coding agents to people who are not engineers.
On this page · 6 sectionsOpen
Internal Goldman Sachs guidance updated in October 2026 limits Claude Code and GitHub coding tools to engineers, Bloomberg News reported on November 2.
Engineers, about 12,000 at the bank, and strats who mix trading, quant research and engineering keep access; most bankers and traders lose it.
Traders had used the agents to monitor client flows, model market moves and positions, and build written and graphic material.
Goldman’s spokesperson cited productivity and risk management; no incident, replacement tool or Anthropic comment has been reported.
In July, CIO Marco Argenti praised vibe coding by traders and bankers, so the October guidance reverses the bank’s own message.
§ 01What Goldman changed
The change is a restriction on coding agents, not a ban on AI at the bank. It narrows who may use two specific tools.
| Group or tool | What changed | Reported by |
|---|---|---|
| Traders and bankers, Claude Code | Access removed | Bloomberg News |
| Traders and bankers, GitHub coding tools | Access removed | Bloomberg News |
| Engineers (about 12,000 at Goldman) | Keep access | Bloomberg News, State of AI |
| Strats (trading, quant research and engineering) | Keep access | Bloomberg News |
| Hong Kong bankers, Claude | Barred earlier, in April 2026 | State of AI |
Carcaterra said the bank makes certain AI tools available to staff “to determine the best use cases” across the firm. According to The Decoder’s summary of the Bloomberg report, he added that the experimentation phase is over and Goldman is now limiting the tools to the people “who are best suited to leverage them for the long term.” Goldman has not named a replacement coding tool for traders and bankers, and Anthropic has made no public comment on the change.
§ 02What traders were using the agents for
The traders who lost access had found real work for them. People familiar with the matter told Bloomberg that the tools were being used to monitor client trading flows, model market moves and positions, and put together written and graphic material. Some traders are frustrated by the change, one of those people said.
The rollback is also a reversal of the bank’s own message. As recently as July, chief information officer Marco Argenti praised “vibe coding” by non-technical staff: a trader would build a first version of a tool with AI and hand it to one of the firm’s engineers to make it production grade.
| When | What happened |
|---|---|
| April 2026 | Goldman bars its Hong Kong bankers from Claude, per State of AI |
| July 2026 | CIO Marco Argenti praises vibe coding by traders and bankers |
| October 2026 | Internal guidance limits Claude Code and GitHub to engineers and strats |
| November 2, 2026 | Bloomberg News reports the change |
| November 8, 2026 | The story reaches the Hacker News front page |
§ 03Why a coding agent is riskier on a trading desk
Goldman has not described an incident, and none has been reported. The stated reason is risk management in general. The specific risks below are our reading, and State of AI makes a similar case.
A coding agent like Claude Code reads files, writes programs and runs them, usually with the same access as the person using it. Code an engineer writes with it normally passes through review, testing and deployment controls. A script a trader builds to watch client flows can slip into daily use without any of that, on a machine that can see client information.
Banks also have to keep information barriers between groups, for example between bankers working on deals and traders who could profit from knowing about them. Software built in a prompt by someone on one side of that wall was not designed with the wall in mind. And on a trading desk, code that mishandles one unusual case can mean real losses for someone who cannot read what the agent wrote.
Cost is the open question. Agent use is billed by the token, and Goldman has not said whether spending played a part; Bloomberg’s sources named productivity as one factor in the review.
§ 04The Anthropic angle
The timing is awkward for both companies. State of AI notes that Goldman is a lead underwriter on Anthropic’s planned listing, so the bank is helping sell shares in the company whose coding agent most of its traders can no longer use. Our running record of the listing is on the Anthropic IPO prospectus page.
Other banks are going the other way. State of AI reports that Barclays expects half its developers to be using Claude Code by the end of the year. Like Goldman, it is giving coding agents to engineers.
§ 05What this means for teams outside banking
Most companies are not bound by information barriers, but Goldman’s question applies everywhere: who owns code an agent wrote for someone who is not an engineer? Three habits make the answer concrete before anything reaches real data.
- Name an owner. Every tool someone outside engineering builds with an agent gets a named reviewer before it touches production data or customers.
- Keep the existing rules. Data access, approvals and change controls still apply when the code came from a prompt.
- Record the use case. If access is pulled, write down what people were using the tool for. Goldman’s traders still need to watch client flows and model positions.
The same rules apply to coding agents’ secrets: Claude Code’s saved permissions recently left 1,122 public GitHub files holding at least one secret, which we covered in Claude Code settings leaking secrets.
§ 06Sources
- Bloomberg News, “Goldman Restricts Traders and Bankers From Using Some AI Tools,” November 2, 2026, via Bloomberg Law
- State of AI, Goldman Sachs Is Pulling Claude Code From Its Bankers and Traders, November 2, 2026
- The Decoder, Goldman Sachs pulls AI coding tools from traders and bankers, November 2, 2026
- Hacker News discussion, November 8, 2026
Q1Who reported the Goldman change first?
Bloomberg News, on November 2, 2026, citing internal guidance updated in October and people familiar with the matter.
Q2Who still has Claude Code at Goldman?
Engineers, about 12,000 at the bank, and strats, the staff who combine trading, quantitative research and software engineering.
Q3What were traders using the tools for?
Monitoring client trading flows, modeling market moves and positions, and putting together written and graphic material, per Bloomberg’s sources.
Q4Did Anthropic comment?
Anthropic has made no public comment on the change. Goldman is a lead underwriter on Anthropic’s planned listing.
Q5What should other companies take from it?
Name an owner for every agent-built tool, keep existing data and change controls, and record what people use a tool for before pulling it.