Don’t let the LLM decide who is allowed to act

Description and use

An automated agent on Google Cloud uses Gemini to analyze resources and applies changes only when cost‑saving thresholds are met, with all critical actions requiring manual approval.

Problem

Automated cloud resource cleanup can unintentionally delete or modify resources if the LLM’s decisions are not constrained by human oversight and precise approval data.

Outcome

The agent reduced manual cleanup effort, prevented silent cost‑draining bugs, and ensured all high‑impact changes required human approval.

Workflow

• 1. Deploy Cloud Run service with write credentials.
• 2. Agent uses Gemini 3.5 Flash‑Lite to read resource fleet and estimate savings.
• 3. Decision logic checks if tool is irreversible or savings ≥40.0 → await approval; if savings ≥5.0 → apply; else skip.
• 4. Agent logs actions, wraps any user‑supplied text in <untrusted> delimiters, and escalates if cost threshold is met.
• 5. Human reviews Level‑2 actions and approves or rejects.
• 6. Agent applies approved changes or skips.
• 7. Periodic audit reports include engine used and decision rationale.

Evidence

The agent cleans the name, wrapped it in <untrusted> delimiters, declared it as data, flagged it to me — and escalated it to Level 2 anyway, because $487 a month is $487 a month no matter what the resource is called.

Source

If the ticket said "1 vCPU and 2Gi" and the executor applied 512Mi. Four code paths had an or "512Mi" fallback, so the ticket text and the applied change were derived independently from the same input.

Source