Agent Signal

The agent signal guardrail lets your AI agents request human review by embedding a special HTML comment in the email body. When Sluice detects the comment, it flags the email for review and shows the agent's message to the reviewer. The comment is automatically stripped before the email is forwarded — the customer never sees it.

How it works

Your AI agent includes an HTML comment in the email body using this format:

<!-- sluice: I'm not confident about the pricing details in this email -->

When Sluice processes the email:

1. The comment is detected and the email is flagged as orange (review recommended)
2. The agent's message ("I'm not confident about the pricing details in this email") is displayed to the reviewer in the dashboard
3. The reviewer can approve, edit, or reject the email
4. The <!-- sluice: ... --> comment is automatically removed before the email is forwarded — the recipient never sees it

Setting up your AI agent

Add instructions to your agent's system prompt telling it when and how to use the sluice comment. Here's an example:

When drafting an email, if you are uncertain about any factual claim, pricing, policy detail, or if the customer's request is ambiguous, include an HTML comment in the format <!-- sluice: your concern here -->. This will route the email to a human reviewer before it's sent. Be specific about what you're uncertain about.

Examples

Uncertain about a claim:

<p>Your Enterprise plan includes up to 50 user seats.</p>
<!-- sluice: I'm not sure if the Enterprise plan limit is 50 or 100 seats. Please verify before sending. -->

Ambiguous customer request:

<p>I'd be happy to process that refund for you.</p>
<!-- sluice: The customer's message was ambiguous — they might be asking for a refund or a credit. Needs human judgment. -->

Edge case the agent can't handle:

<p>I've escalated your request to our team.</p>
<!-- sluice: This customer is asking about our service in a country we don't operate in. I'm not sure how to respond accurately. -->

Use cases

Building safety into your prompts — Instead of hoping your AI agent is always right, give it a mechanism to say "I'm not sure." This is much safer than having the agent guess and potentially send incorrect information.

Gradual autonomy — Start with broad agent signal usage ("flag anything you're less than 90% confident about"), then narrow it over time as you build trust in your agent's responses.

Training data — The flagged emails and agent comments become valuable training data for improving your agent's prompts and knowledge base over time.