Write Your First Policy

Behavry uses Open Policy Agent (OPA) with Rego policies to decide what agents can and cannot do. Every tool call is evaluated in real-time.

Policy decisions

Every policy evaluation returns one of three decisions:

Decision	Effect
Allow	Tool call is forwarded to the target server
Deny	Tool call is blocked; agent receives an error
Escalate	Tool call is held in a queue for human approval

Policies are default deny — if no rule explicitly allows an action, it is blocked.

---

Your first policy

Create a file at policies/base/my-first-policy.rego:

package behavry.authz

import future.keywords.if

# Allow all actions for the developer role
allow if {
    input.agent.role == "developer"
    input.action.type in {"read", "list"}
}

# Deny write actions to production paths
deny if {
    input.action.type == "write"
    startswith(input.resource.path, "/prod/")
}

# Escalate file deletions
escalate if {
    input.action.type == "delete"
}

Activate the policy

curl -X POST http://localhost:8000/api/v1/policies \
  -H "Authorization: Bearer $ADMIN_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
    "name": "my-first-policy",
    "description": "Basic developer restrictions",
    "content": "'"$(cat policies/base/my-first-policy.rego)"'"
  }'

The policy is evaluated on every tool call within seconds of activation.

---

Testing the policy

# This should succeed (allowed)
# curl ... your agent making a read call

# This should fail (denied)
# curl ... your agent writing to /prod/

# This should trigger escalation (visible in dashboard)
# curl ... your agent deleting a file

Open the dashboard → Escalations to approve or deny the escalation.

---

Policy input shape

Every policy receives this input object:

{
  "agent": {
    "id": "agent-uuid",
    "name": "my-agent",
    "role": "developer",
    "risk_tier": "low"
  },
  "action": {
    "type": "write",
    "tool": "filesystem/write_file"
  },
  "resource": {
    "path": "/prod/config.yaml",
    "server": "filesystem"
  },
  "session": {
    "id": "session-uuid",
    "duration_seconds": 120
  },
  "requester": {
    "id": "user@company.com",
    "channel": "claude-code"
  }
}

→ Full Policy Engine docs