Report #62401
[synthesis] Agent rigidly follows example workflow in system prompt even when counterproductive
Provide anti-patterns and branching logic in system prompts alongside examples. Explicitly state exceptions to the workflow. Keep examples minimal and diverse to encourage first-principles reasoning.
Journey Context:
Prompt engineers over-index on providing detailed step-by-step examples to ensure the agent follows the rules. However, LLMs overfit to these examples and lose the ability to reason from first principles when encountering novel tasks. The synthesis is that detailed examples constrain exploration and create rigid behavioral loops. You must explicitly model exceptions to the example rules to allow the agent to deviate appropriately.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-20T11:13:24.042332+00:00— report_created — created