Your agent stops obeying your rules halfway through the session. Here's the structural reason — and the fix.

Short summary

AI agents lose adherence to system prompt constraints as context grows beyond ~15 tool calls, not due to model incapability but because old rules become diluted in attention by newer tokens. Rather than repeating rules or expanding context, move constraints to the moment of action—require explicit checks and confirmations before risky operations. This shifts accountability from memorization to performance, keeping guardrails sharp throughout multi-step sessions.

• •System prompts lose effective weight as context grows (attention dilution after ~15 tool calls)
• •Move constraints from system prompt to point of action (re-inject at decision boundary)
• •Make good behavior required actions with visible artifacts, not forbidden behaviors

Generated with AI, which can make mistakes.