Govern · Policy Engine
Available nowTurn AI policy into code you can version
Halo policy bundles are structured, versioned rules with allow, deny, and alert outcomes. Run a rule in audit mode to observe, switch to enforce mode when you are ready, and read the explanation behind a decision.
- modes, audit to observe and enforce to act
- 2
- structured authoring against a schema
- YAML
Why it matters
A policy document does not stop anything
Acceptable use rules for AI usually live in a PDF or a wiki page. Nothing connects those words to the tool calls and model requests that agents actually make.
- 01
Rules that live outside the system depend on people remembering them.
- 02
Without versions, nobody can say which rules applied when something happened.
- 03
Switching on a strict rule without testing it first can block legitimate work.
How it works
From draft to enforced policy
Each stage uses the same bundle, so what you test is what you enforce.
- 1
Draft a bundle
Write rules in structured YAML that follows the policy schema. A rule has a category, conditions, an outcome, a priority, and a mode.
- 2
Run it in audit mode
Assign the bundle with rules in audit mode. Halo records what each rule would have done while actions continue.
- 3
Review the evidence
Check the matches in the evidence trail, adjust conditions and priorities, and confirm that each rule hits only what you intend.
- 4
Enforce and version
Switch the rule to enforce mode. Bundle versions change when rules change, so past decisions stay tied to the rules that governed them.
Capabilities
Policy that behaves like code
Bundles are structured, versioned, and testable, so changes can be reviewed, rolled out gradually, and traced afterward.
Versioned policy bundles
Bundles carry a version and a lifecycle of draft, active, and archived, so you can see which rules governed a past run.
Allow, deny, and alert
Rules return allow, deny, or alert outcomes, and a rule can also hold an action for approval, so one policy can cover routine, sensitive, and prohibited actions.
Audit and enforce modes
Audit mode records what a rule would have done and lets the action proceed. Enforce mode applies the outcome. You choose the mode for each rule.
Tool-level checks
Conditions can match the tool, its arguments, the destination, the agent, and the time, so a rule can be as narrow as one tool call.
Model request and response checks
Rules can evaluate model requests and responses, including classifier results from supported gateway paths, and respond with the outcome you define.
Decision explanations
Decisions record which rule matched and why, so a reviewer can explain an outcome without reverse-engineering the policy.
Safe rollout
Audit first, enforce when the evidence agrees
A new rule can start in audit mode, where Halo records the decision it would have made and lets the action continue. Reviewers read those records, fix conditions that match too much, and enforce only after the matches look right.
- Audit mode records would-have-blocked decisions
- Priority order decides which rule applies first
- Deny wins when rules at equal priority conflict
- Switching a rule to enforce needs no redeploy
FAQ
Questions, answered
Policies are structured bundles that follow a schema and can be written as YAML. A rule names its category, conditions, outcome, priority, and mode, which makes policies easy to review, version, and compare, the same way you handle other code.
The action is allowed by default. Teams that want an allow-list posture add a low-priority catch-all deny rule. Evidence from audit mode shows what your agents actually do, which helps you decide what the catch-all should cover.
Rules are evaluated in priority order, and the first matching rule applies. If two rules at the same priority match, deny takes precedence, which is a conservative default that keeps an overlapping allow from overriding a deny.
Halo evaluates policy at its model and tool gateways for agent traffic, and supported browsers evaluate browser policy locally in audit, warn, or block mode. The pages for each surface describe what that surface checks.
No software can do that on its own. Halo records evidence that supports governance programs aligned to the EU AI Act, NIST AI RMF, and ISO/IEC 42001. It does not automate compliance mapping, and your organization remains responsible for its decisions and obligations.
Put a halo around the AI your business runs.
See discovery, runtime policy, approvals, and evidence working together on your own use cases.