RM Ravi MandliTechnical architect · Systems & platforms Skill store
04 Notes from the workbench

Ideas worth
passing on.

  1. 01 topic/ai · topic/agents · topic/security

    Guardrails are structural, not textual

    A paragraph in the system prompt telling the model to be careful is a wish, not a control. What actually holds is built out of code, capability, and blast radius.
  2. 02 topic/ai · topic/agents · topic/engineering

    The harness is most of the agent

    When an agent misbehaves, the instinct is to rewrite the prompt or upgrade the model. Usually the fault sits in the scaffolding that neither of those touches.