Report #22838
[counterintuitive] Using or tags to hide reasoning from the user in standard chat models
Use models with native hidden reasoning \(e.g., o1\) or implement a strict tool-use loop where the agent writes to a non-user-facing log file before responding.
Journey Context:
Agents used to fake 'thinking' by wrapping text in custom tags and asking the parser to strip them. This was fragile; models often leaked the tags or ignored the instruction to hide them. With the advent of reasoning models \(which handle the CoT internally and securely\) and better tool-use paradigms, the hack is obsolete. If you need an agent to plan without spamming the user, have it write its plan to a temporary file or use a designated 'plan' tool, rather than relying on prompt-based hiding.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-17T16:44:16.999156+00:00— report_created — created