WEBVTT

00:00:00.000 --> 00:00:01.580
An AI agent approval workflow

00:00:01.580 --> 00:00:03.320
should classify the next action,

00:00:03.480 --> 00:00:06.400
not rely on the agent's name or a vague feeling of trust.

00:00:08.840 --> 00:00:10.720
A practical matrix uses four lanes.

00:00:13.560 --> 00:00:14.700
First, observe and prepare.

00:00:15.080 --> 00:00:16.620
The agent reads permitted sources

00:00:16.620 --> 00:00:18.260
and drafts local analysis.

00:00:18.860 --> 00:00:21.980
Second, propose a change: it names the exact target,

00:00:22.140 --> 00:00:24.200
evidence, expected effect, and rollback,

00:00:24.660 --> 00:00:25.840
but has no write capability.

00:00:28.210 --> 00:00:30.370
Third, apply a bounded reversible change

00:00:30.370 --> 00:00:32.850
only after approval is tied to the operation,

00:00:32.850 --> 00:00:34.510
target, and reviewed revision.

00:00:35.310 --> 00:00:37.010
Fourth, escalate for a human decision

00:00:37.010 --> 00:00:39.770
when the effect expands into publication, spending,

00:00:40.270 --> 00:00:43.270
permissions, deletion, or external communication.

00:00:44.450 --> 00:00:45.710
Useful approval is specific.

00:00:46.270 --> 00:00:47.790
It identifies who can approve,

00:00:48.190 --> 00:00:49.410
what operation is allowed,

00:00:49.790 --> 00:00:51.810
the exact target and payload version,

00:00:52.310 --> 00:00:53.290
the supporting evidence,

00:00:53.830 --> 00:00:54.910
when approval expires,

00:00:55.390 --> 00:00:57.530
and how the result will be checked or reversed.

00:00:58.590 --> 00:01:00.430
Technical controls should enforce those limits.

00:01:01.110 --> 00:01:02.770
Separate read and write access,

00:01:03.030 --> 00:01:04.790
require exact identifiers,

00:01:05.090 --> 00:01:06.690
preview changes where possible,

00:01:07.070 --> 00:01:08.350
prevent duplicate retries,

00:01:08.650 --> 00:01:10.010
and record the after-state.

00:01:10.870 --> 00:01:12.330
Then test awkward cases:

00:01:12.690 --> 00:01:13.350
missing data,

00:01:13.690 --> 00:01:14.690
ambiguous targets,

00:01:15.070 --> 00:01:15.990
changed proposals,

00:01:16.370 --> 00:01:17.450
and partial failures.

00:01:18.170 --> 00:01:20.590
This matrix is an editorial planning model.

00:01:20.970 --> 00:01:22.770
It is not evidence that it has been deployed

00:01:22.770 --> 00:01:24.390
or tested in Daniel's systems.