Goldman Sachs didn't have a Claude problem

Last month Goldman Sachs updated its internal guidance so that Claude Code and GitHub's coding tools are now only available to its engineers. Traders and bankers who had been using them to follow client flows, model positions and build their own tools have lost access.

Photo by Jakub Żerdzicki on Unsplash

What I find interesting is the timing. In July, Goldman's CIO was talking publicly about the tools traders had been building with AI. Three months later, those same tools are gone.

The easy conclusion is that AI coding agents are too risky for people who aren't engineers. I don't think that's the right one.

If you look at the reasons that were reported, none of them are really about the model. The agent runs with the same access as the person using it. Code a trader builds doesn't go through the review and deployment checks that engineering code does. Goldman built information barriers, between deal teams and trading desks, into its own AI platform, but a tool someone puts together in an afternoon doesn't have them. And the person running the code doesn't always understand what it's doing.

You could swap Claude for any other agent tomorrow and you'd have exactly the same problems. The real issue is that nothing sits between the agent and what it's able to do, and nobody is clearly responsible for it.

First of all, banks built their controls for two things: people, and approved software. An agent is somewhere in the middle. It gets every permission the person using it has, it can build things that never go near the deployment pipeline, and the audit trail shows who was logged in rather than what the agent actually did. So the controls are all still there, the agent just goes around them.

Taking the tools away doesn't really solve this. The traders still need to watch client flows and model positions, so that work goes back into the engineering queue, or into spreadsheets. Goldman has more than 12,000 engineers, so it can absorb it. Most regulated firms can't. For a specialist bank or a building society, "only engineers" in practice means nobody.

This is exactly the problem we're building Senni for.

Senni is the runtime AI agents execute through in regulated firms. Agents don't run alongside it, they run through it, which means every action goes through Senni before it happens. In a situation like Goldman's, that means:

• Every agent, and every tool it uses, has a named owner. A tool a trader builds can't touch production data until someone has taken responsibility for it.
• The firm's existing rules are checked on every single action. If a trader's agent tries to read data from the deal side, it gets blocked, even if the trader's own laptop could reach it.
• Everything the agent does is recorded in an audit log that can't be altered, so when compliance asks what happened and who approved it, the answer is already there.
• Senni runs inside the firm's own cloud, so the data never leaves.

I want to be honest about what Senni doesn't do: it doesn't make the code right. If a trader's model handles an edge case badly, it's still wrong. What we make sure of is that it can only do what it's allowed to do, that someone owns it, and that there's a record of everything it did. Without that, you can't even start looking for the mistake.

When something goes wrong, the regulator won't ask which model you used. They'll ask who was responsible. Goldman's answer was to limit who can use agents. I think the better answer is to let the people closest to the problem use them, and make sure every action goes through controls you can actually prove.

If you're working through this at your firm, I'd be keen to hear how you're approaching it. senni.io