Intent Verifier
Your agent drifts.
We catch it.
GPT-4o-mini checks every action against the session's original intent. When the agent starts doing something different from what it was told, we flag it - before it causes damage.
The Problem
An agent told to 'send a $500 invoice to Ahmed' might - due to a bad lookup, a prompt injection, or model drift - send $5,000 to Sara instead. Intent Verifier catches the gap between what you asked and what the agent is about to do.
How it works
1. Session created with intent - natural language description of goal.
2. Action evaluated - GPT-4o-mini scores action against intent.
3. Drift score calculated - 0.0 (perfect match) to 1.0 (mismatch).
4. Threshold flagged - action paused, human notified, audit log updated.
Intent Verifier SequenceEvaluating
Original Intent
"Send invoice to Ahmed for $500"
Agent Target - Step 2
"stripe.invoices.create (Recipient: Sara, Amount: $5,000)"
Drift Coefficient0.92 / 1.0
Threshold exceeded (0.70) - session suspended