Applied AI
3 notes
RAG Is Not a Mandatory Pipeline, but an On-Demand Evidence Capability
Understand the goal and inspect the current source first; call Local RAG for evidence candidates only when cross-source information can genuinely improve the judgment.
From Workflow Harness to Lightweight Prompts: What I Removed After Codex 5.6
When a model's native workflow becomes capable enough, a Harness can shift from an efficiency multiplier to a context burden. This rewrite keeps the safety boundaries while substantially simplifying the Prompts and runtime.
Only Four Hard Rules at the Top Level: Where an Agent Must Stop
Models can become increasingly autonomous, but production, persistent data, costs, and irreversible actions must still require explicit authorization. Safety is not the opposite of efficiency; it is the prerequisite for automation.