How It Works

1 Policy Engine

Your rules, enforced automatically. Define policies in plain English like "Allow all read operations" or "Deny access to production databases" or "Require approval for payments over $1,000".

Actions matching your policies are handled instantly - allowed or denied in under 50ms with no human intervention needed. The vast majority of agent actions are handled at this layer.

Policy outcomes:

• ALLOW - Action proceeds immediately
• DENY - Action is blocked with reason
• APPROVAL_REQUIRED - Action queued for review

2 AI Consensus Council

When an action doesn't match any explicit policy, it enters the Consensus Council - a panel of independent AI judges that evaluate the request against your organization's intent and risk profile.

Each judge operates independently, analyzing:

• The action being requested and its parameters
• The context and history of the agent session
• Your organization's policies and stated intent
• Potential risks and edge cases

Council thresholds:

• 80%+ agreement to approve - Action is auto-approved
• 80%+ agreement to reject - Action is auto-denied
• 50%+ disagreement - Escalate to human oversight

This handles the long tail of edge cases without flooding your team with approval requests.

3 Human Oversight

Only truly ambiguous decisions reach your team. When AI judges disagree (50%+ split), the action is escalated for human review.

What you see:

• Full context of the action request
• Each judge's reasoning and vote
• Risk assessment and relevant policies
• One-click approve or deny

In practice, this means you're reviewing the genuinely hard cases, not rubber-stamping routine operations. Most organizations see less than 5% of actions reaching human review.