Goldman gave traders AI agents. Then took them back.

WHAT HAPPENED

Goldman limited coding agents to its engineers

Per Bloomberg, October guidance restricts Claude Code and GitHub to engineers. Traders and bankers had used them to track client flows and model markets. In July, the CIO was still praising trader-built tools.

THE NUMBER

86%

of tasks had a compliant path, yet top models still overstepped

Frontier models drifted past their authorization at rates of 0.43 to 0.61, with no adversary involved.

THE REAL PROBLEM IN COMPANIES

What I see in companies: tools ship before owners exist

1. 01Someone outside engineering builds a script, and it quietly reaches daily use.
2. 02Nobody is named as responsible for what the agent wrote or what it touched.
3. 03Permissions match the user, not the task, so the agent can reach far more than it needs.

I'VE ALREADY SOLVED THIS

I run this in production for an enterprise account

1. 01A multi-agent platform in production at a US$5.5 million account.
2. 02Multi-agent in production requires an audited defect log.
3. 03About 60% efficiency gain, with the governance built in.

MY TAKE

Launching is not the same as being used.

Handing out access is easy. Scoped permissions, audit trails and a clear owner are what make it a deployment. In practice, governance designed in from day one is cheaper than rolling a tool back after something breaks.