WEBVTT

00:00:01.000 --> 00:00:04.970
An AI agent can choose next steps, but its authority needs

00:00:05.000 --> 00:00:06.490
clear boundaries first.

00:00:07.160 --> 00:00:10.650
A fictional company wants two software packages connected.

00:00:10.840 --> 00:00:14.890
A project assistant may look up customer details, identify missing

00:00:14.920 --> 00:00:17.730
information and prepare a follow-up question.

00:00:17.960 --> 00:00:21.690
It may not commit to a quotation or send a customer message.

00:00:21.960 --> 00:00:25.850
Describe the information the agent receives, the tools it may choose

00:00:26.120 --> 00:00:27.930
and when its work is complete.

00:00:28.200 --> 00:00:31.850
Here, a finished draft is the endpoint, rather than a project

00:00:31.880 --> 00:00:34.135
sold without human involvement.

00:00:34.160 --> 00:00:37.655
The application checks every requested system action.

00:00:37.680 --> 00:00:41.770
Telling a prompt not to send messages is not access control.

00:00:41.800 --> 00:00:46.410
Give the integration only the permissions required for this particular task.

00:00:46.440 --> 00:00:49.975
If the sequence is always fixed, an ordinary workflow may be

00:00:50.000 --> 00:00:51.050
simpler.

00:00:51.080 --> 00:00:54.890
An agent becomes worth investigating when the next step genuinely

00:00:55.080 --> 00:00:58.170
depends on information found during the task.

00:00:58.200 --> 00:01:01.530
If a source is missing or an instruction requests a forbidden

00:01:01.560 --> 00:01:05.050
action, the agent stops or asks for review.

00:01:05.320 --> 00:01:09.930
Bound time, cost and action counts too, so an unclear assignment

00:01:10.040 --> 00:01:11.450
cannot run indefinitely.

00:01:12.040 --> 00:01:16.170
Try normal requests, missing information and messages that attempt to

00:01:16.200 --> 00:01:17.490
change the task.

00:01:17.720 --> 00:01:21.290
Check that forbidden actions are blocked and that proposals reach the

00:01:21.320 --> 00:01:23.930
designated reviewer for an explicit decision.

00:01:24.520 --> 00:01:28.810
Choose one task, representative examples and a responsible staff

00:01:28.840 --> 00:01:29.290
member.

00:01:29.800 --> 00:01:35.530
Assess the output and actions taken before adding tools or granting wider authority.

00:01:35.880 --> 00:01:41.130
Read OXIAE’s guide to building AI agents with explicit permissions,

00:01:41.160 --> 00:01:43.290
tests and human oversight.