Policy Engine

The policy engine controls what your AI agents are allowed to do. Policies are written in YAML and evaluated in order — the first matching rule wins.

Policy structure

version: "1"
rules:
  - tool: github          # tool name or "*" for any
    action: DELETE        # action/HTTP method or "*" for any
    decision: deny        # allow | deny
    message: "DELETE is blocked"

  - tool: "*"
    action: POST
    require_reasoning: true
    message: "POST calls must include reasoning"

  - tool: stripe
    action: charge
    decision: allow

Rule fields

Field	Required	Description
tool	yes	Tool name or * to match any
action	yes	Action name, HTTP method, or *
decision	no	allow or deny (default: allow)
require_reasoning	no	If true, call is denied unless reasoning is provided
message	no	Human-readable message returned on denial
id	no	Optional rule name for audit log references

Matching logic

Rules are evaluated top-to-bottom. The first rule where both tool and action match is applied. If no rule matches, the call is allowed by default.

rules:
  - tool: github
    action: DELETE
    decision: deny          # ← matches first for github+DELETE

  - tool: "*"
    action: DELETE
    decision: deny          # ← never reached for github+DELETE

Wildcard matching

• tool: "*" matches any tool
• action: "*" matches any action
• Exact strings are case-insensitive

Requiring reasoning

require_reasoning: true means the call is only allowed if the reasoning field (or X-Agent-Reasoning header) is non-empty.

- tool: stripe
  action: "*"
  require_reasoning: true
  message: "All Stripe calls require agent reasoning"

If reasoning is missing, the call is denied with the specified message.

Example policies

Allowlist only (deny everything else)

version: "1"
rules:
  - tool: github
    action: GET
    decision: allow

  - tool: "*"
    action: "*"
    decision: deny
    message: "Only GitHub GET calls are permitted"

Read-only (block all writes)

version: "1"
rules:
  - tool: "*"
    action: POST
    decision: deny
    message: "Write operations are not permitted"

  - tool: "*"
    action: PUT
    decision: deny

  - tool: "*"
    action: DELETE
    decision: deny

Require reasoning for sensitive tools

version: "1"
rules:
  - tool: stripe
    action: "*"
    require_reasoning: true

  - tool: database
    action: "*"
    require_reasoning: true

Managing policies

Policies can be created and edited in the Policies tab of the dashboard, or via the API:

curl -X POST https://api.agentic-guard.com/agents/{agent_id}/policy \
  -H "Authorization: Bearer $JWT" \
  -H "Content-Type: application/json" \
  -d '{"rules_yaml": "version: \"1\"\nrules:\n  - tool: \"*\"\n    action: DELETE\n    decision: deny"}'

Policy hot-reload

Policies are reloaded on every request — no restart required.