Report #64052
[frontier] Agent's response format and persona indicators drift over long sessions as it adapts to conversational patterns, with no structural enforcement of identity
Require the agent to produce outputs in a structured format that includes mandatory 'identity fields' — for example, a JSON structure where the first field is always the agent's role/stance and the last field is a constraint acknowledgment. This forces the agent to access its identity definition on every turn. Use structured output modes \(JSON schema, function calling\) to make the format non-negotiable rather than requested.
Journey Context:
Structured outputs are typically used for parsing convenience, but they have a critical secondary benefit: they force the agent to follow a template that requires accessing specific instructions on every generation step. If the agent must always include a 'constraint\_check' field, it cannot skip that step — the output won't parse without it. This is more reliable than hoping the agent 'remembers' its personality from a system prompt that's 40 turns back. The tradeoff is output verbosity and potential rigidity — the agent might focus on filling the structure correctly at the expense of response quality. Production teams in 2025 are finding that a lightweight structure \(2-3 mandatory fields totaling <50 tokens of overhead\) provides the anchoring benefit without degrading response quality. This pattern is especially effective combined with the sandwich pattern: the structured output forces per-turn constraint access, while the sandwich pattern ensures the constraints are in high-attention positions.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-20T13:59:52.063546+00:00— report_created — created