WEBVTT

00:00:00.250 --> 00:00:03.066
Agents ask their models a lot of questions.

00:00:03.512 --> 00:00:06.211
Sometimes, a good answer already exists.

00:00:06.241 --> 00:00:08.904
A shim sits between your agent and its model.

00:00:08.934 --> 00:00:11.704
It's built to give cheap but valuable direction,

00:00:11.779 --> 00:00:14.940
or a direct response with a reusable component.

00:00:14.970 --> 00:00:16.043
Here's how it works.

00:00:16.073 --> 00:00:17.494
A request arrives.

00:00:17.524 --> 00:00:21.246
Before any model runs, the shim checks its saved rules.

00:00:21.437 --> 00:00:25.505
Today, we saw them match by keyword, or by a simple word list.

00:00:25.535 --> 00:00:29.152
It's also built to match the meaning, with a small classifier.

00:00:29.190 --> 00:00:31.077
Then one of three things happens.

00:00:31.268 --> 00:00:33.049
First, a direct response.

00:00:33.079 --> 00:00:36.545
The shim answers by itself, and no model is called.

00:00:36.736 --> 00:00:38.007
Second, direction.

00:00:38.037 --> 00:00:41.282
The shim adds a short note, and then the model runs.

00:00:41.845 --> 00:00:43.616
Third, no match.

00:00:43.749 --> 00:00:46.576
The request passes through, unchanged.

00:00:46.640 --> 00:00:48.723
This is the real, live preview.

00:00:48.753 --> 00:00:51.464
It needs no key, and it calls no model.

00:00:51.597 --> 00:00:53.866
We ask: what is a Task Object?

00:00:53.896 --> 00:00:57.432
A saved rule matches, and the shim answers directly.

00:00:57.565 --> 00:01:00.358
No model call, so no model tokens.

00:01:00.430 --> 00:01:03.664
Now we ask for a flyer for our nonprofit food bank.

00:01:03.925 --> 00:01:06.798
The word list catches the word nonprofit.

00:01:06.932 --> 00:01:10.165
So the shim adds a note of about seven hundred characters.

00:01:10.240 --> 00:01:15.251
It says that crisis work stays with people, and that a different reviewer checks the work.

00:01:15.511 --> 00:01:18.141
Then the model starts with that direction.

00:01:18.180 --> 00:01:19.694
Now, the honest part.

00:01:19.724 --> 00:01:23.258
We tried ten everyday prompts today, and three matched.

00:01:23.288 --> 00:01:25.303
Only one got a direct answer.

00:01:25.436 --> 00:01:27.242
The other seven passed through.

00:01:27.317 --> 00:01:29.657
An earlier check matched one of six.

00:01:29.848 --> 00:01:33.628
Say food bank without the word nonprofit, and nothing matches.

00:01:33.761 --> 00:01:36.657
So today, the shim knows a small number of things.

00:01:36.687 --> 00:01:39.143
Teaching it more is the work ahead.

00:01:39.180 --> 00:01:41.160
Direction isn't free, either.

00:01:41.351 --> 00:01:43.702
A note adds tokens to the request.

00:01:44.056 --> 00:01:48.625
So a shim must earn its place, with notes that are short, and only when they help.

00:01:49.188 --> 00:01:52.060
We haven't measured savings from these notes yet.

00:01:52.090 --> 00:01:55.405
Any agent can try the preview today, with no key.

00:01:55.852 --> 00:01:59.109
Send it a prompt, and see what the built-in rules would add.

00:01:59.369 --> 00:02:03.938
To learn how we train Meta models for this work, open Fabric Intelligence.

00:02:04.501 --> 00:02:07.619
Connect your agents at teleionlabs.com.