Jev Tool-Call Guardrails for AI Agents | AutoJev

Jev tool-call guardrails add a structured review before an AI agent performs a consequential action. The guard evaluates the proposed tool call, its side effects, existing safeguards, policy thresholds and reversibility, then recommends allow, confirm, review or deny.

The workflow question

Can this agent perform the proposed action now, or does it need confirmation or review?

The recommendation does not grant authority. Your application must still enforce authentication, authorization, approval thresholds and tool-specific validation.

Example guard input

{
  "tool": "issue_customer_refund",
  "action": "Refund USD 680 after a disputed duplicate charge",
  "arguments_summary": ["order_id=ord_7429", "amount_usd=680"],
  "side_effects": ["Moves funds", "Changes the order payment state"],
  "safeguards": ["Customer identity and duplicate charge verified"],
  "policy": ["Refunds above USD 500 require human approval"],
  "reversibility": "partially_reversible"
}

An illustrative result shape is:

{
  "recommendation": "review",
  "confidence": 0.91,
  "risk_score": 0.78,
  "guidance": "Request human approval before calling the refund tool."
}

Actual answers and probabilities depend on the supplied state and configured Jev provider.

Good guardrail boundaries

• Sending email or messages to external recipients.
• Issuing refunds, payments or credits.
• Deleting, overwriting or publishing data.
• Deploying production code or changing infrastructure.
• Running shell commands with irreversible effects.
• Modifying permissions, accounts or security settings.

Low-risk read-only calls usually do not need the same review depth. Apply the guard where consequences justify the added decision.

Run the preset

Use POST /api/v1/decisions/tool-guard or open the tool-guard example in the Playground. The MCP tool is autojev_guard_tool_call, and the matching Skill teaches an agent when to invoke it.

Related Jev resources

• Jev AI for agent guardrails
• Configure the Jev MCP server
• Install the tool-guard Skill
• All Jev use cases

Frequently asked questions

Does AutoJev block the tool call itself?

No. The preset returns a typed recommendation. Enforcement remains in your application or an explicitly configured hook.

Can a guardrail replace authorization?

No. Authorization answers whether the caller has permission. The Jev guard evaluates semantic risk and policy context within that existing permission boundary.

Should secrets be included in the request?

No. Send a compact summary of arguments and side effects. Do not send passwords, provider keys or unrelated private data.