Partly, and you should keep using them. Rules files are real institutional memory for agents. Here's what they can't do:
Rules files
AgentTrail
Rules filesAdvisory.An instruction the model may ignore, and instruction-following decays over long sessions.
AgentTrailDeterministic.The destructive command doesn't run, whatever the model decided. Enforcement lives outside the model.
Rules filesHand-written.Someone has to notice the third repeat and write the rule.
AgentTrailDetected.The pattern engine finds the repeats in your sessions and suggests the policy.
Rules filesFragmented.Per-repo, per-tool: CLAUDE.md doesn't govern Cursor; .cursor/rules doesn't govern Claude Code. They drift.
AgentTrailOrg-wide.One versioned corpus, every repo, every developer.
Rules filesUnmeasured.A markdown file can't tell you whether the rule worked.
AgentTrailMeasured.Violations per policy per week, before/after repeat rates, backtests.
Rules filesNo audit standing.Compliance can't cite a repo file as a control.
AgentTrailEvidence.Every policy decision is an exportable OTel span.
And the two work together. Keep CLAUDE.md and .cursor/rules as your soft-guidance layer; AgentTrail turns the repeats it finds into enforced policies, and reserves hard enforcement for the things that must never happen.