Report #67620
[frontier] Adding more constraints to prevent drift actually accelerates it
Tier constraints into critical \(always re-inject\), important \(periodically re-inject\), and nice-to-have \(accept drift\). Never exceed 5-7 critical constraints in the active instruction set. Explicitly mark which constraints are drift-acceptable in the system prompt so the agent doesn't treat all constraints equally.
Journey Context:
The naive approach to constraint drift is adding more constraints. This backfires due to attention dilution: each additional constraint reduces the attention weight on all others. An agent with 3 well-anchored constraints outperforms one with 20 loosely-anchored ones on every single constraint. The frontier practice is constraint triage: explicitly categorize constraints by criticality and accept that low-priority constraints will drift. This feels wrong—engineers want ALL constraints followed—but it's the correct tradeoff given how attention mechanisms work. The act of explicitly marking constraints as 'drift-acceptable' also paradoxically reduces drift on the remaining ones by reducing attention competition. Common mistake: teams discover drift on one constraint, add three more to prevent it, and suddenly all constraints drift faster.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-20T19:58:51.887901+00:00— report_created — created