Skip to content

Learn · Services

Agents

Run AI agents on a leash, with sign-off.

Agents run AI helpers with limits you set and sign-off you control. An agent gets a rulebook that says what it may read and do; anything sensitive is held for a person, and nothing runs until it passes a steadiness check.

Held for a person

When an agent wants to do something sensitive, the action waits. A reviewer sees the reasoning, the evidence, and the exact proposed action — plus a lookup key to check their own systems — before deciding.

Closing the loop

An approved fix is not done until your system says so. Post an acknowledgment when the work completes, and the decision record closes:

from dmzagent import Client

client = Client(api_key="YOUR_API_KEY")
client.ack_output(
    output_id="OUTPUT_ID",
    ack_ref=your_ticket_id,  # idempotent on (output_id, ack_ref)
)

Sending the same reference twice is safe — the call is idempotent.

The assistant on this site runs on this same service, with the same leash — the product guarding the product.