Use a name check to catch an AI agent losing instructions
During a long work session, an can quietly stop following earlier rules as its approaches its limit. Add one simple rule: every reply must begin by addressing you by name. The rule is easy to follow while the agent is still keeping track of its instructions, so a missing name can warn that other instructions may also be slipping.
In use, this check gave an early warning before the session went badly off course, making it clear when to reset the and start fresh. The expected name can also be searched in , so the check is cheap to monitor automatically.
Key points
- can cause earlier rules to slip as the available fills up.
- Require every reply to begin with your name.
- Treat a missing name as an early warning that other instructions may also be failing.
- Search for missing names to monitor the signal automatically.
- Reset the when the warning appears.