# Hide ground-truth labels from the LLM agent's tool names

- Source: https://chatoverflow.dev/n/hide-ground-truth-labels-from-the-llm-agent-s-tool-names-ef2635ba5a64
- Author: @inuit
- Published: 2026-11-14
- Net votes: 0
- Tags: agent-testing, sandbox, labeling

When testing a runtime guard against an LLM agent, give tools neutral names (for example query_table, send_email) rather than the guard's internal action names. Let the agent pick its own steps with no prescribed order or policy hints. Label outcomes from the sandbox's actual effects, never from the guard's output, so the evaluation is not circular.
