Skip to content

Set standing rules for agents

Applies to OML v0.6.x · Status: Stable · Last verified 2026-11-06

A guardrail is a reviewed rule every agent session loads at its start, such as "never push to the release branch" or "run the tests before every commit". It is a memory of kind guardrail. Agents or people write it, and it applies only after a different identity approves it.

  • Where it applies. From the agent's workarea and, as an administrator sets on its key, the workareas above it. Each rule is shown with where it was set and who approved it. On the web, Guardrails lets you choose a workarea and shows whether the key reads that workarea alone or its readable parents too.
  • Loaded at the start. The plugin's session-start hook puts the rules in front of the agent and again after a compaction and in helper agents.
  • Checked at the tool call. A rule with a rule body is checked before each matching tool call and refuses or asks; it never allows.
  • Severity. never, must and should.
  • Without the plugin's hooks. oml guardrails export --format agents-md|claude-md --out <path> writes the promoted set as a file, labeled "text, not enforcement; do not edit". oml guardrails show prints what a session would load.

What a rule cannot do. A rule the agent only reads is an instruction to a model, not a guarantee. Only the tool-call check binds. An exported file is a copy that withdrawal and erasure do not reach. We have not measured whether agents follow rules they only read.