Agents can have intents and goals, and their behaviour should be consistent with their overall intent. i.e., an agent should not be surprised by its past behaviour even if it conflicts with the current short-term goals.

We assume the intent is encoded in the system prompt by the Principle of prompt identity