Oleksandr Kuzmenko, EditorPublished 6 min readAI-assisted · editor-reviewed
TL;DR
- Autonomous CLI agents like Claude Code require explicit sandbox permissions and cannot be treated like passive chat interfaces.
- Regulated environments require a designated software engineer to review agentic scripts before they process production or sensitive client data.
- Enterprise unmetered agent access leads to rapid token inflation, prompting firms to enforce hard per-seat budget limits.
Key facts
- Engineering workforce
- 12,000 of 47,000 staff
- Accenture Claude subsidy
- $29 per employee monthly
- Policy update date
- October
The End of Unguided Vibe Coding on Trading Desks
Internal guidance updated at Goldman Sachs restricts Anthropic's Claude Code and Microsoft's GitHub to the bank's roughly 12,000 engineers, ending self-directed agent use among traders and bankers. While desk strats who bridge quantitative research and engineering retain access, general staff can no longer run agentic tools. Previously, non-technical personnel used Claude Code to monitor client trading flows, assemble presentation materials, and model market positions directly on machines with sensitive client data.
Information Barriers and Execution Risks
Unlike passive code completion, Claude Code reads the local filesystem, writes full programs, and executes commands autonomously with the permissions of the active user session. In highly regulated banking environments, financial institutions must maintain strict information barriers between deal-making bankers and market traders. When employees generate and run unreviewed scripts, these firewalls erode. Furthermore, traders lacking software engineering training cannot inspect how generated scripts handle edge cases, posing substantial financial risk.
Token Budgets and Governance Guardrails
Token consumption also influenced the policy reassessment. Enterprise-wide agent rollouts frequently encounter token inflation; Amazon previously terminated an internal leaderboard when employees ran agents on frivolous tasks, and Accenture capped Claude subsidies at $29 per employee monthly. Goldman's decision indicates that enterprise AI strategy is shifting from ubiquitous end-user vibe coding to structured, engineer-governed deployments where named software engineers maintain formal code review and testing accountability.
When to use
- When technical developers need an agent to read repository files, run terminal commands, and draft PRs under review.
- When quant engineers or 'strats' work within sandboxed testing environments with automated CI/CD safeguards.
When NOT to use
- When non-technical personnel run autonomous code-executing agents on local workstations holding confidential client records.
- When scripts generated by prompt-based vibe coding are deployed directly into live operations without engineer signoff.
What to do today
- Audit developer and non-technical machine permissions to ensure CLI agents cannot execute bash scripts without review.
- Enforce pre-commit hooks and pull request gates for any script generated by autonomous coding tools.
- Set hard token spending quotas at the team or organizational API gateway level.
Step names are taken from What to do today.
Step names are taken from What to do today. Audit developer and non-technical machine permissions to ensure CLI agents cannot execute bash scripts without review.. Enforce pre-commit hooks and pull request gates for any script generated by autonomous coding tools.. Set hard token spending quotas at the team or organizational API gateway level.
- Audit developer and non-technical machine permissions to ensure CLI agents cannot execute bash scripts without review.
- Enforce pre-commit hooks and pull request gates for any script generated by autonomous coding tools.
- Set hard token spending quotas at the team or organizational API gateway level.
What the community says
It’s the case in all industries. People have been feeding medical patient data into Claude
This level of reaction only comes when someone steps in shit. So, there likely was some large leak or “mistake” that has the potential to get out and this is part of either covering up or reducing how bad