Goldman Sachs Is Pulling Claude Code from Its Bankers and Traders

Goldman Sachs updated internal policy guidance in October to restrict Anthropic's Claude Code and Microsoft's GitHub exclusively to its software engineers, cutting off access for non-technical traders and investment bankers across the firm. The bank initiated this policy change after reviewing internal productivity and operational risk concerns surrounding autonomous third-party software agents that write and execute code independently on client-facing machines. As detailed in thestateofai.com's report, the decision effectively ends a period where front-office personnel directly deployed AI-generated scripts to track client trading flows and construct market models.

This article details the structural factors behind the policy change, the operational boundaries established between technical and non-technical staff, and the distinct governance models currently adopted across major financial institutions. It provides financial technology leaders, quantitative analysts, and compliance officers with an overview of how Wall Street is managing information barriers, software review requirements, and the deployment of autonomous coding systems.

What shipped in Goldman Sachs Is Pulling Claude Code from Its

The October policy shift formalizes a strict organizational boundary that confines Claude Code and GitHub access to Goldman Sachs's dedicated engineering division, which comprises more than 12,000 of the institution's roughly 47,000 total global workforce. Quantitative strategists, known internally as strats who operate across trading desks, quantitative research, and software engineering, also retain operational access to these developer utilities. For the rest of the firm's trading desks and banking divisions, third-party autonomous agents capable of generating and executing arbitrary code have been disabled, leaving those employees without direct access to prompt-driven terminal assistants.

To maintain baseline generative capabilities across the company, Goldman Sachs continues to provide its proprietary, multi-model in-house AI assistant to all employees. Unlike third-party developer agents, this internal assistant operates within firm-wide information barriers and does not execute autonomous local machine commands. Front-office traders who previously relied on Claude Code to monitor live client flow data, generate market position models, and assemble written materials must now route new software requests through formal engineering channels. Goldman Sachs spokesperson Nick Carcaterra noted that the institution distributes select artificial intelligence systems across teams specifically to determine optimal operational use cases while balancing regulatory requirements.

What improved in Goldman Sachs Is Pulling Claude Code from Its

The restriction establishes enforceable controls over software deployment and eliminates uninspected code execution on workstations handling sensitive client transactions. Previously, scripts assembled by non-technical staff skipped standard code review pipelines, automated unit testing, and mandatory compliance validation before touching market positions. Confining agent access to engineers ensures that enterprise software development aligns with regulatory information barrier rules, preventing cross-desk data exposure between investment bankers managing corporate transactions and traders handling secondary markets. The shift also curtails runaway token expenses caused by unrestricted employee prompt loops, replacing unmetered individual experimentation with governed developer infrastructure.

By standardizing agent utilization inside software engineering, the bank eliminates situations where critical trading algorithms operate without developer oversight. Trading operations avoid unexpected position losses stemming from unhandled runtime edge cases in scripts generated by non-programmers. Meanwhile, core back-office engineering projects remain active, including systems developed during a six-month onsite engagement with Anthropic engineers to automate trade accounting and client onboarding pipelines.

What you gain from Goldman Sachs Is Pulling Claude Code from Its

Financial enterprises and technical organizations gain a clear blueprint for separating developer productivity tooling from end-user workstation access. Operating through a unified engineering tier ensures that every automated script interacting with client data adheres to existing data entitlements, compliance mandates, and audit trails. Engineering teams reclaim complete visibility over the codebase, preventing shadow scripts from becoming critical points of operational failure on active trading desks. This separation allows compliance teams to satisfy regulatory constraints while preserving developer productivity gains for the staff trained to audit, debug, and maintain complex code.

Organizations also gain predictable budget boundaries and operational clarity regarding who owns automated code generation. By clarifying that software creation requires professional software testing, management avoids situations where non-programmers ship opaque code that cannot be fully diagnosed during production outages. Technical departments can systematically collect the automation needs identified by frontline traders, converting ad-hoc terminal scripts into professionally maintained services that integrate seamlessly with backend trading systems.

How to get Goldman Sachs Is Pulling Claude Code from Its

Authorized developers managing Claude Code installations can update the client, switch runtime configurations, and manage model targets using standard terminal interfaces. To update an existing global installation to the latest build, execute the package manager upgrade command:

npm install -g @anthropic-ai/claude-code

Alternatively, running the built-in update tool from within the application directory refreshes the runtime binaries:

claude update

When operating inside the interactive terminal environment, developers switch target models by invoking the in-app model selection command:

/model fable

To initialize a new session pinned to a specific model version directly from the command-line interface, launch the executable with the model flag:

claude --model fable

To configure the chosen model as the persistent default for all future terminal sessions, export the primary environment variable within your shell configuration:

export ANTHROPIC_MODEL=fable

What to watch after Goldman Sachs Is Pulling Claude Code from Its

The primary event to monitor is Anthropic's upcoming initial public offering, expected to list on the Nasdaq under the ticker ANT before the end of November, following an S-1 registration statement filed on October 26. Goldman Sachs and Morgan Stanley serve as the lead underwriters for the listing, placing Goldman bankers in the position of pitching shares for a company whose flagship coding product they are restricted from using. Industry analysts are also watching whether Anthropic can defend its projected valuation targets following recent market volatility and reported investor pushback over broader artificial intelligence infrastructure capital expenditures.

Disparate approaches across competitor institutions will provide valuable benchmarks for enterprise AI governance. While Barclays announced plans for half of its developers to adopt Claude Code by year-end alongside roughly 5,000 corporate job reductions, lenders like JPMorgan and Charles Schwab have focused on barring external autonomous agents from interacting with customer accounts. Furthermore, geographic and regulatory hurdles remain active considerations, evidenced by Goldman Sachs and JPMorgan removing Claude access for Hong Kong personnel earlier this year over contract boundaries.