The Pullback Is a Requirements Document
The headline that made the rounds last week was “Goldman Sachs is pulling Claude Code from its bankers and traders.” It collected the usual two reactions. The skeptics declared the AI coding experiment over: a serious institution finally admitted the stuff doesn’t work. The boosters declared Goldman a dinosaur: a compliance culture strangling its own productivity while nimbler firms eat its lunch.
Both camps skipped the detail that makes the story interesting. Goldman’s engineers kept their access. So did the Strats, the quantitative hybrids who live between trading and software. Per Bloomberg’s reporting, the October guidance removed Claude Code and GitHub from bankers and traders specifically, following what the bank described as a review of productivity and risk concerns. This came just months after Goldman’s own CIO was publicly praising the tools traders had built for themselves.
So the accurate version of the story is this: a bank made powerful code-generation tools available to tens of thousands of people whose job is not software, watched what happened for a while, and then drew a line that happens to trace the border of its engineering controls. That is not a verdict on AI. That is a verdict on deploying a capability ahead of its control structure. The pullback is not a retreat. It is a requirements document, written in the only language an organization can write one after the fact.
Honesty requires naming the rival explanation before arguing with it. Goldman’s own framing was productivity and risk, and the boring reading is available: thousands of non-engineers burning tokens on tools that did not pay for themselves, and a bank doing cost-benefit rather than compliance philosophy. Probably both are true. But notice that the two readings converge on the same line. Engineers and Strats are where the productivity is provable, and they are also where the controls already exist. Those are not two different facts. Code that flows through review, pipelines, and ownership is code whose value an institution can see, measure, and keep. The uncontrolled path did not just carry risk; it made the benefit illegible. Either way you read the review, the conclusion is the same sentence: bring the capability onto the controlled path.
Shadow IT at machine speed
Every large company has fought shadow IT for decades: the spreadsheet that became load-bearing, the Access database running a desk’s risk process, the macro nobody can explain whose author left in 2019. The industry built entire governance functions to contain it. The containment mostly worked because shadow IT was slow. A motivated trader could build one dangerous spreadsheet a quarter.
Coding agents removed the speed limit. Give a banker Claude Code and you have not given them a productivity tool, you have given them a software development shop with no review process, no deployment pipeline, no maintenance plan, and no institutional memory. The tools traders built, the ones the CIO praised in July, were real and probably useful. They were also unreviewed code, written by people who could not review it themselves, running against live workflows, inside one of the most heavily regulated environments on earth.
I spent years as a security architect in fintech before I ever ran an engineering org, and the thing that struck me about the reporting was how familiar the shape is. Nothing about this problem is new. What is new is the rate. Shadow IT used to accrete. Now it compiles.
What a bank actually worries about
No incident has been reported, and the honest reading is that Goldman did not need one. You do not wait for the crash to decide the brakes are required. Here is what the risk review almost certainly saw, because it is what anyone who has worked inside one of these perimeters would see:
Unreviewed code in load-bearing positions. Not “bad code.” Unreviewed code. The distinction matters. Engineering organizations do not review code because engineers write bad code; they review it because review is how an institution takes ownership of a change. A tool a trader built with an agent has no second set of eyes anywhere in its history. When it misprices something, the question “who understood this?” has an uncomfortable answer.
Information barriers. Banks maintain legally mandated walls between departments: between those who advise on deals and those who trade on the market, between public-side and private-side information. The entitlement systems that enforce those walls were built to gate what a human can access, and an agent inherits its user’s entitlements, so the naive fear of an agent reading what its user cannot is mostly handled. The actual problem is movement. Agents industrialize exactly the flows the entitlement layer was never built to watch: context assembled from a dozen sources, material pasted in to be helpful, generated connectors that quietly bridge systems. The wall gates access. It has never had to police movement at this rate. My day job is legal tech, where the walls between matters are load-bearing in the same way, and I can tell you that making an agent respect a confidentiality boundary is a genuine engineering problem, not a checkbox.
Comprehension debt. The trader who built the tool rotates desks. The tool keeps running. Nobody on the desk can read it, including the person who made it, because making it did not require reading it. Engineering organizations have a word for software with no owner who understands it, and the word is incident, future tense.
Attribution. When something goes wrong, a regulated institution must be able to say who did what, when, and under whose authority. “An agent did it at the direction of someone in a non-engineering role, outside the deployment pipeline” is not an answer any compliance officer wants to write down.
None of this requires believing the models write bad code. Every item on the list holds even if the generated code is excellent. That is what the booster camp keeps missing: these objections are not about quality, and better models answer none of them.
The part Goldman got right, and the part everyone else should steal
Banks already know how to manage dangerous automated things. They have been doing it for decades with trading models. There is an entire discipline, model risk management, with Federal Reserve guidance behind it, built on the premise that a model whose behavior you cannot fully explain must be inventoried, validated, monitored, and owned by someone accountable. The failure here was not that Goldman lacked a framework. It is that coding agents walked in the front door as “productivity software” and never got routed through it.
Notice what the restriction actually preserved. The reporting does not say why the line sits where it sits, but it sits somewhere precise: engineers and Strats, the people whose output already flows through review, CI, staged deployment, and on-call ownership. Whatever the internal reasoning, the line lands exactly where the controls are, and the effect is a policy worth stating out loud: code enters this institution through the controlled path or it does not enter. That policy is correct, and it was correct before the first LLM was trained.
So read the pullback as the requirements document it is. For agents to be handed back to the bankers and traders, somebody has to build:
- Attribution by construction. Every agent action traceable to a human sponsor with the authority to request it.
- Reconstructable sessions. Logs sufficient to replay what the agent saw, what it did, and what left the boundary. Not for debugging. For the regulator.
- Barrier-aware context. The agent’s retrieval and memory respect the same information walls the humans do, enforced in the platform, not the prompt.
- A controlled path for citizen-built tools. A pipeline where a desk-built tool gets review proportional to its blast radius, an owner of record, and a retirement date, before it touches a live workflow.
- Secrets and egress handled by the platform. Assume anything in the agent’s context can leave it. Nothing should depend on the agent’s discretion, because the agent does not have any.
None of that is exotic. It is change management, rewritten for a change author that is not a person. Which has been my argument about AI in high-stakes systems all along: the model is not accountable and was never going to be. When the agent-built tool misfires, the human who deployed it is the one on the hook, and an institution that cannot name that human has a governance gap, not an AI problem.
The requirements are permanent
The pullback is temporary. Goldman is not abandoning agents; it is re-sequencing, controls before capability, which is the only order that was ever going to survive contact with a regulator. The April removal in Hong Kong, the October guidance, these are an institution learning in public what every regulated adopter will have to learn somewhere: the perimeter does not move for the tool. The tool grows up to meet the perimeter.
The firms sneering at Goldman this week will rediscover each line of that requirements document one incident at a time. The ones worth watching are the firms quietly building the controlled path right now, because the prize at the end is not modest. It is every banker and trader building their own tools again, except this time with an audit trail. The tension between velocity and governance is real; anyone who has waited three weeks on a change advisory board knows it. But it is a design input, not a law of nature. Shipped in the right order, controls first, you keep most of both. Shipped in the wrong order, you eventually keep neither.