Goldman Sachs Pulls AI Coding Tools From Traders and Bankers: What Is Confirmed

Goldman Sachs has reportedly pulled advanced AI coding tools from its traders and bankers, according to Bloomberg's report as summarized by The Decoder. Tools such as Anthropic's Claude Code and Microsoft's GitHub tools are now limited to engineers. The bank says the experimentation phase is over.

This is a policy shift at one of the most closely watched banks on Wall Street. It reverses a "let non-engineers build" stance that the bank's own CIO praised only months ago. Below we separate what is confirmed from what is not, add the history, and list what a team outside banking can take from it.

What is confirmed about the Goldman Sachs AI coding restriction

The core facts come from Bloomberg, which reported the restriction on November 2, 2026. The Decoder summarizes it as follows:

• Most employees, including traders and bankers, lost access to advanced AI coding tools.
• The tools named are Anthropic's Claude Code and Microsoft's GitHub tools.
• Only engineers can still use them.
• The change comes from internal guidelines dated October.
• Traders and bankers had used the tools for months to monitor trading flows and to create content.
• The rollback reportedly frustrates staff.

Goldman spokesperson Nick Carcaterra gave the reasons only as a "variety of factors including productivity and risk management." He said the experimentation phase is over and that the tools will be limited to people "who are best suited to leverage them for the long term."

We could not open the Bloomberg article itself, which is paywalled. This post relies on The Decoder's summary of it and on the spokesperson's quoted statement. Where details are missing, we say so.

What did Goldman's CIO say about vibe coding in July?

Per The Decoder, Goldman CIO Marco Argenti praised vibe coding by non-technical staff in July. The model was simple. A trader or an asset management employee builds a quick AI prototype of a tool. An engineer then takes it and scales it for production.

That model is attractive for a bank. Traders know exactly what a monitoring screen should show. They rarely have engineering time on their desk. AI coding agents close that gap, and a prototype that works on live data answers a question faster than a ticket to the technology team.

The reported October guidelines end that plan for anyone who is not an engineer. The prototype step moves back inside engineering. Goldman has not said how the hand-off will work now, or whether a controlled environment will replace the open access.

Why would a bank pull coding agents from non-engineers?

Goldman named "productivity and risk management" and nothing more specific. We do not know the internal reasoning. We can lay out the risks that banks and other regulated firms commonly weigh with agentic coding tools. These are explainx.ai's analysis, not Goldman's stated reasons.

Access scopes for AI tools shown as cables passing a ring, a view of AI coding tools governance at a bank

Code that nobody reviews

A coding agent writes working code quickly. The author is often not trained to read it. In a trading or banking setting, code that touches positions, client data or market feeds needs review. When the author cannot review it, the review burden moves to engineers, who then own code they did not write. A prototype can quietly turn into a production dependency.

Data and system access

Claude Code and similar tools run commands, read files and call APIs from a user's machine. A trader's machine can reach sensitive market and client data. Without a sandbox and a strict permission model, an agent with broad access is a larger attack surface. We covered the broader pattern in Claude Code security guidance, and the risk shows up in real incidents such as a contractor running a personal Claude Code agent on a Defense Department network.

Productivity that is hard to measure

Goldman lists "productivity" first. After months of experimentation, a bank can ask a hard question: did the prototypes save time, or did they create work for engineers? The bank has not published numbers either way.

Regulatory and records obligations

Banks must keep records, supervise communications and control model risk. Content made with AI, and tools built outside normal change control, create gaps in those processes. This is a general point about the sector. Goldman did not cite any regulator.

Is this part of a pattern at banks?

Banks have moved in both directions at once over the past year. They push AI into employee workflows, and they tighten the edges.

• In April 2026, Bloomberg reported that Goldman staff in Hong Kong lost access to Anthropic's Claude. That was a regional restriction and a separate matter from the October guidelines.
• In October 2026, major US banks moved against AI agents acting on customer accounts. See our coverage of Citi, Bank of America and JPMorgan blocking AI agents and of banks detecting unauthorized AI agents.
• Banks also use AI heavily inside engineering. Bank of America put 1,000 contractors on patching flaws that Anthropic's Mythos found.
• On the labor side, Goldman's own economists estimated that AI could displace 9% of workers, and Barclays tied 5,000 job cuts to AI.

The picture is not "banks reject AI." Banks keep AI where engineers and controls surround it, and they pull it back where unreviewed output could reach production, clients or regulators.

What does "experimentation phase is over" mean in practice?

The phrase is the most useful sentence in the coverage. It frames the earlier open access as a pilot, and it signals a move to a managed program.

A prototype model on a drafting triangle, a picture of the vibe coding prototype stage that now moves into engineering

A pilot lets many people try a tool to find out what works. A managed rollout picks the use cases that worked, chooses the people who run them, and sets rules for data, review and audit. Goldman's wording, "people who are best suited to leverage them for the long term," points at the second model.

Two readings are possible, and the public facts do not decide between them:

1. A real retreat. The tools caused enough risk or low-quality output that the bank cut access.
2. A planned consolidation. The bank learned what works and is moving the tools into engineering, with a possible controlled path for desk staff later.

The spokesperson's wording leans toward the second reading, and staff frustration suggests the change felt like a retreat from the user side. We label this as interpretation. Goldman has not said which one is true.

What can other teams learn from Goldman's reported change?

You do not need to be a bank to use these points.

1. Put an end date on pilots. Say when open experimentation stops and who decides what comes next.
2. Write the access rules before the habit forms. It is easy to give a tool to everyone. It is hard to take it back, as the reported staff frustration shows.
3. Keep the prototype-to-production hand-off explicit. If non-engineers prototype, define who reviews, who owns the result and what data the prototype may touch.
4. Measure productivity honestly. Count review time and rework, not only time saved writing the first draft.
5. Limit what an agent can reach. Use sandboxes, scoped credentials and logging. For agent-specific controls, AgentBeam is the agent security platform from the explainx.ai team that stops AI agents before they take dangerous actions.

If you manage a team that uses AI coding agents, compare your own rules against these five points. Our guide to Claude Code security guidance is a good place to start.

What is still unknown?

• The full text of the October guidelines.
• How many employees lost access, and in which divisions.
• Whether other AI tools, such as chat assistants, are affected. The report names coding tools.
• Whether a controlled path will exist for desk staff, for example a sandboxed internal environment.
• Whether the change relates to any incident, audit finding or regulator request. Goldman named none.
• Whether the bank's contracts with Anthropic or Microsoft change.

We will update this post if Goldman or Bloomberg publishes more. If you work at a bank and have a primary document, the details above are what readers most want.

Why this matters beyond Wall Street

Coding agents spread fast because they work. They also spread to people who were never trained to review code. That gap is the real issue in this story, and it exists in every company that handed out a coding agent in the past year. Goldman, a firm with deep engineering resources, still concluded that unrestricted access for non-engineers did not fit its long-term plan.

For builders, the lesson is practical. A coding agent in the hands of a domain expert is a strong prototyping tool. The same agent without review, sandboxing and ownership is a source of hidden dependencies. Decide which one you are running.

Related reading

• What is vibe coding? A 2026 explainer
• Claude Code security guidance plugin
• Citi, Bank of America and JPMorgan block AI agents on customer accounts
• Banks detect unauthorized AI agents on customer accounts
• Goldman Sachs: AI could displace 9% of workers
• Barclays 5,000 job cuts and AI
• Accenture contractor's personal Claude Code agent on a DoD network
• SUSE cuts 10% of staff, citing automation and AI

Details are accurate as of November 2, 2026. The Bloomberg report is paywalled and the guidelines are not public, so facts may change.