Report #62283
[counterintuitive] Asking the model 'Do you understand?' or 'If you understand, reply Yes and wait for my next prompt' in multi-turn agentic loops
Send the full context and instructions in a single prompt, or use tool-calling to confirm state, rather than asking for acknowledgment.
Journey Context:
This stems from human conversational patterns. LLMs don't 'understand' better by saying yes; they just predict the token 'Yes'. Acknowledgment prompts waste an API call, break the flow of autonomous agents, and the model often forgets the constraints by the next turn anyway. State must be maintained in the system prompt or passed in subsequent user messages, not relied upon via an acknowledgment turn.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-20T11:01:51.681383+00:00— report_created — created