OVERT-as-Code quickstart — AI governance as code
This quickstart takes you from an empty file to a validated, tested AI governance policy you can run in CI.
Prerequisites
Section titled “Prerequisites”- Rust 1.75+ (2021 edition) — for the preview build from source
- A terminal
-
Write a policy. Create
overt.toml:[policy]id = "my-agent-v1"name = "My Agent Governance Policy"enforcement_mode = "shadow" # observe first; switch to "enforce" when readyprofile = "enterprise-general"[tool.defaults]mode = "deny"[[tool.allow]]name = "search_docs"classification = "read_only"[[tool.allow]]name = "send_email"requires_approval = true # human-in-the-loop gateapproval_timeout_secs = 120[[tool.deny]]name = "delete_*"reason = "Destructive operations require a separate workflow" -
Validate it. Catch errors before they reach production:
Terminal window glacis overt validate overt.tomlValidation complete: 0 error(s), 0 warning(s)Policy 'my-agent-v1' v1.0.0 is VALIDPolicy hash: f2f14dfac40cb97e -
Write a test suite. Assert decisions for specific requests in
tests.toml:name = "My Agent Policy Tests"[[tests]]name = "destructive tool is denied"[tests.request]type = "tool_call"tool_name = "delete_account"[tests.expect]decision = "deny"reason_contains = "Destructive" -
Run the tests:
Terminal window glacis overt test overt.toml -t tests.tomlPASS: destructive tool is denied1 passed, 0 failed -
Wire it into CI. Fail the build on policy regressions:
- run: glacis overt validate overt.toml- run: glacis overt test overt.toml -t tests.toml
Next steps
Section titled “Next steps”- Move from
shadowtoenforceonce you trust the decisions. - Export OSCAL for your compliance team.
- Understand how this differs from OPA/Terraform.
- When you need proof that the policy ran — signed, verifiable receipts — see the runtime product and the verifier.