October 9, 2026
A limit in an agent's prompt doesn't always hold: open-source CSL-Core checks tool calls before they run
A limit in the prompt saying “do not return more than $500” holds “most of the time” and leaves no trace when it breaks. The open CSL-Core, described on Oct 9, resolves each call before the function runs, and the model never sees the rules.
A limit in the prompt saying “do not return more than $500” holds “most of the time” and leaves no trace when it breaks. The open CSL-Core, described on Oct 9, resolves each call before the function runs, and the model never sees the rules.
The author advises tweaking the strict default rules on a busy agent.