Agent Drift Is a Config Problem, Not a Model Problem

Most teams blame model randomness when their AI agent behaves inconsistently. Wrong diagnosis. Agent drift happens when the SOUL.md (or equivalent config) is too vague to constrain behavior across edge cases. The model fills the gaps with reasonable-sounding behavior that's wrong for your context. The Test Read your agent config cold. Can a stranger predict exactly what the agent will never do? If not, you have a scope problem — not a model problem. What Vague Looks Like role : " You are a helpful assistant. Help users with their tasks." What edge cases does this cover? None. The agent will invent behavior for every ambiguous situation it encounters. What Tight Looks Like role : " You are a data validation agent for financial records." constraints : - Never modify source files directly - Write validated output to /output/validated/ only - If record fails validation, append to failed-records.jsonl with reason - Stop and escalate if failure rate exceeds 5% of batch - Never infer missing