Goldman Sachs Restricts Claude Code and GitHub Access to Engineers Only

Goldman Sachs has restricted advanced AI coding tools, including Anthropic's Claude Code and Microsoft's GitHub, to its engineers. The change was set out in internal guidelines from October, according to Bloomberg, as summarized by The Decoder on 2 November.

Traders and bankers had spent months using the same tools to monitor trading flows and draft content. They no longer can.

The reversal is notable because Goldman's own CIO made the opposite argument in public only a few months ago. In July, CIO Marco Argenti praised "vibe coding" by non-technical staff in an interview, describing a model where a trader or someone in asset management would build a quick prototype with AI and hand it to an engineer for production. The Decoder reports that is no longer the plan.

What Goldman says changed

The explanation offered so far is thin. A Goldman spokesperson, Nick Carcaterra, cited "a variety of factors including productivity and risk management," according to The Decoder's account of the Bloomberg report. He added that the experimentation phase is over and the firm is now limiting the tools to the people "who are best suited to leverage them for the long term." That is the only on-record framing available from the dossier: productivity and risk, no further detail on which risk or whose productivity.

Read plainly, the two factors pull in opposite directions. Productivity argues for spreading the tools as widely as possible. Risk management argues for narrowing them. Goldman appears to have decided the second outweighs the first, at least for staff whose job is not writing software, and the reported internal frustration suggests not everyone inside agrees with the call.

The move also lands in a wider pattern. Several outlets in the dossier's recent coverage treat enterprise AI coding access as a governance question rather than a procurement one, and the infrastructure being built around agents suggests the same conclusion.

The plumbing is getting serious

Uber's approach shows what governed agent access looks like at scale. InfoQ reported on 2 November that Uber built an MCP Gateway to standardize how AI agents discover and invoke internal services, supporting more than 800 MCP servers and 5,000 tools. The design splits a registry, which holds tool definitions, schemas, ownership and enablement settings, from a proxy gateway that translates MCP requests into HTTP, gRPC or TChannel calls. A crawler called AutoCrawler scans internal interface definition language registries, extracts API definitions from Protocol Buffers and Thrift schemas, and generates MCP-compatible schemas, using a large language model to write the tool descriptions.

The detail that matters for Goldman-style decisions is the default. According to InfoQ, newly discovered tools stay disabled until their owning teams review and explicitly enable them. Authorization is enforced at both server and tool level through Uber's internal access control system, and the gateway can redact personally identifiable information from tool responses. Uber also built Omni MCP for incremental discovery, because loading definitions for thousands of tools into an agent's context is inefficient. That is the control layer a bank needs before it lets a trading desk point an agent at internal systems.

The agent vendors keep pushing autonomy

While Goldman tightened, the toolmakers kept loosening. DeepSeek previewed a plan for a more autonomous agent that handles multiple tasks with minimal human intervention, according to the South China Morning Post on 1 November. The system is dubbed a super agent, and the paper was published by the company's DeepSeek Harness team on Saturday. The stated aim is to reduce the need for humans to coordinate multiple agents and monitor their progress, which the company frames as a bottleneck in complex workflows.

Google went the other way on control, in a sense: it put the choice of agent in the developer's hands. In a developer blog post dated 29 October, Google said Android Studio now supports more than 40 ACP-compatible coding agents through Bring Your Own Agent and the open Agent Client Protocol, stable in the Android Studio Rabbit 2 release. Google also shipped Change Review, which generates a report of potential issues in agent-made changes and can have the agent fix what it finds, plus Android Bench 2.0, a leaderboard evaluating models and agent harnesses on long-running Android development tasks. The pitch is flexibility; the subtext is that reviewing what an agent did is now the hard part.

That is the tension Goldman is navigating. The tools are good enough that non-engineers want them, and the vendors are making them more autonomous, not less.

Developers are being asked to trust less

The scepticism is not confined to banks. Meta's engineering blog published results of a Python Typing Survey 2026 on 28 October, examining AI tools and the checker landscape. EE Times ran a design data management piece on 3 November arguing that as digital workflows move into text-based environments like VS Code, engineers still lose time switching between editors, command-line tools and separate repositories to manage versions and design history. Neither is a direct attack on AI coding, but both describe a workflow where the surrounding tooling, not the model, decides whether work goes well.

NVIDIA's developer blog made the case for simulation on 29 October, describing how developers turn ideas into simulations with frontier AI agents in its Omniverse series. NVIDIA also published on 28 October about RTX Spark Windows laptops gaining new models, agent tools and games, and on 22 October about Cosmos opening tools to train, evaluate and deploy world models. The common thread across the vendor posts is that agents need environments, evaluation harnesses and guardrails before they are trusted with anything that matters.

Hardware constraints have not gone away either. Tom's Hardware reported on 3 November that NTDEV, the developer behind Tiny11, booted Windows 11 26H2 on a system with 128MB of RAM, about 3.1% of Microsoft's 4GB minimum, reaching the desktop, File Explorer and Notepad. NTDEV called the result "painfully slow" and treated it as an experiment. The parallel is not exact, but the lesson is: stripped-down systems do remarkable things until someone needs to rely on them.

Elsewhere in the stack

LG CNS completed a 30MW AI data center in Jakarta, Light Reading reported on 2 November, and plans to expand AI data center development into Malaysia, Thailand, Vietnam, the Philippines and Japan. The first phase is an 11-storey facility with 46,281 square meters of floor area and capacity for more than 100,000 servers, with power systems handling more than 300 kilowatts per rack. LG CNS says it leased 100% of the facility before completion. Compute supply is not the constraint Goldman is managing, but it is the reason the tools exist at all.

Robotics is running the same playbook. Agility Robotics and Churchill Capital Corp XI now expect about $465 million in Agility's proposed SPAC merger, The Robot Report reported on 30 October, down from the $620 million predicted at a $2.5 billion valuation in June. The company expects to close on 4 November and says it has more than $300 million in committed multi-year orders for Digit 5, with deliveries to early-access customers in the first half of 2027. The Robot Report's October roundup also noted Schneider Electric's agreement to acquire PTC for about $22.6 billion and Boston Dynamics' new four-fingered hand with 13 degrees of freedom.

None of that changes what happened at Goldman. It does show that the appetite for automation is not shrinking, only the appetite for ungoverned automation.

What to watch

The open questions are concrete. Whether other banks follow Goldman's restriction, and whether the productivity argument or the risk argument wins inside them, will show up in hiring and in tool budgets before it shows up in press releases. Whether Claude Code and GitHub stay engineer-only at Goldman past the current guidelines is also unconfirmed; the dossier contains no statement that the restriction is permanent.

For developers, the near-term signal is that the review layer is becoming the product. Google's Change Review, Uber's disabled-by-default tool registry and Goldman's retreat all point the same way: the assistant writes, but someone still has to sign off. The dossier does not say how Goldman's engineers will handle that sign-off, or whether the traders who lost access will get a narrower tool instead. Asked for detail beyond the spokesperson's statement, the dossier offers none.