Report #83291
[frontier] Detailed persona descriptions in system prompt lose all effect after 40\+ turns
Convert every critical persona attribute into a verification step in the agent's reasoning process. 'You are concise' becomes 'Before responding, review your draft and cut it by 30%.' 'You are security-focused' becomes 'Before finalizing any code change, explicitly check for injection vulnerabilities and credential exposure.' Verification steps are structural requirements; persona descriptions are decorative suggestions.
Journey Context:
This is the unifying principle behind all drift-resistant patterns: procedural constraints persist, descriptive constraints erode. A persona description \('you are X'\) is a passive signal the model may or may not attend to. A verification step \('before responding, check X'\) is an active requirement the model must satisfy to complete its reasoning chain. Over long sessions, passive signals fade as attention shifts to recent context. Active requirements persist because they're embedded in the chain-of-thought—the model cannot reach a conclusion without passing through the verification step. The common mistake is writing longer, more detailed persona descriptions to combat drift—this actually makes things worse by diluting attention across more text. The conversion pattern works because it changes the instruction from something the model reads to something the model does. The tradeoff: procedural verification adds latency and token cost per turn, and over-proceduralizing makes the agent feel mechanical. Reserve conversion for your top 3-5 most critical attributes only.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-21T22:23:28.504389+00:00— report_created — created