Three-tier boundaries is a pattern for writing the constraints in an AI coding agent’s specification as three distinct tiers rather than a flat list of rules: Always do — actions the agent takes without asking, such as running tests before commits; Ask first — actions needing human approval, such as modifying a database schema or adding a dependency; and Never do — hard stops, such as committing secrets or editing vendored directories.
Usage
The pattern is reported to come from a GitHub analysis of over 2,500 agent configuration files, which found it more effective than a simple list of prohibitions because it lets an agent proceed confidently on “Always” items, flag “Ask first” items for review, and hard-stop on “Never” items, rather than treating every rule as carrying the same weight. “Never commit secrets” is reported as the single most common helpful constraint found in that study.
When It Applies
It applies wherever a specification needs to give an agent enough autonomy to act without constant supervision while still gating actions that carry real risk, such as changing CI/CD configuration or database schemas. The source frames it as more nuanced than a flat list of rules precisely because it distinguishes actions that are always safe from ones that need oversight and ones that are categorically off-limits.