0:02 Hi, my name is Elijah Strait and I am
0:04 super excited to be able to demo for you
0:06 today the Microsoft Agent Framework. The
0:09 Microsoft Agent Framework is a new
0:11 open-source engine for building and
0:14 orchestrating intelligent AI agents.
0:15 Today we're going to be looking at a
0:17 clothing company called Zava. Zava
0:20 receives a lot of different pitches for
0:22 various clothing lines that they want to
0:23 people want them to incorporate into
0:26 their overall product line. And this is
0:27 tedious to have to be able to go through
0:29 all of these different pitch decks for
0:31 these various types of close. And so
0:32 we're going to see how the agent
0:35 framework workflows are able to help
0:37 Zava automate this process and reduce
0:40 toil. What's going to happen is first
0:42 the pitch decks are going to be uh
0:44 brought in using a Python script in a
0:46 Python package. What's great about the
0:49 workflow uh architecture is that you're
0:51 able to mix both traditional business
0:54 processes with multi- aent solutions.
0:56 And so from there, after extracting all
0:58 of the information from the pitch deck,
1:00 that information is going to go to a
1:02 subworkflow of concurrent agents that
1:04 are going to do market research, fashion
1:06 design research, as well as production
1:09 feasibility research. The output of
1:10 those agents is then going to go to an
1:13 aggregator. From there, the a report is
1:16 going to be drafted. A human designer is
1:17 going to be able to look at the report
1:19 and see if this is a product line they
1:21 want to incorporate or not. And then a
1:24 final report is then drafted. So today
1:26 we will be actually looking at an
1:28 extremely bad pitch deck called the
1:30 Arctic Blast Coats heavy summer
1:33 collection. Now this is obviously not a
1:35 great collection. Um it's 95 degrees in
1:38 the summertime. Nobody wants coats. Not
1:40 have they don't have a very good target
1:43 customer. Um the product line is not
1:45 awesome. Sales projections are not
1:48 great. And they're really just like not
1:51 a very good company or or product. So,
1:54 let's see if our
1:57 workflow can identify the problems with
1:58 this. While that's running, we're going
2:02 to go over here to VS Code and check out
2:04 how we're combining both our traditional
2:06 business processes with our agents. And
2:08 so, as I mentioned earlier, we're just
2:11 going to extract um the the content of
2:13 these PowerPoint decks using a Python
2:15 script and a package. So, you can see
2:17 here, we're going to validate that um
2:20 the file exists as a PowerPoint file.
2:22 load said file, initialize a data
2:24 structure for getting the the
2:26 information out, process each slide,
2:28 making sure we're getting um text out
2:29 from each shape in the slide, and then
2:32 identifying some keywords. Then from
2:33 there, after we've done that, we're
2:36 going to initialize our agents. And so
2:37 you can see here, we're creating an
2:38 agent with the system prompt, your
2:41 senior fashion analysis research at Zava
2:44 or excuse me, market uh research anal
2:46 analyst at Zava. Made sure to that the
2:48 output is very concise. We don't want it
2:50 to be too verbose for us. And we're
2:53 going to be actually using the Azure AI
2:55 Foundry uh agent service in order to do
2:58 this using the project client. Um and
3:00 then you can see we're creating some
3:01 other agents here as well. And so then
3:03 how do we get these to all come
3:05 together? We're going to be using the
3:08 workflow builder. And so in workflows we
3:11 have these concepts of uh executors and
3:12 edges, which is very similar to nodes
3:14 and edges if you're familiar with
3:16 directed graphs. And we're going to
3:19 start with uh and the beginning executor
3:21 then add a bunch of edges connecting it.
3:22 So you can see here we have our
3:24 concurrent analysis subworkflow which is
3:26 our agents. We also have our report
3:29 writer for writing our report. We have a
3:31 approval manager for the human approval
3:33 step. And then you can see here we also
3:35 have edges with conditional these are
3:37 conditional edges that depending on what
3:40 the human says uh either approves or
3:43 rejects it then sends it to a one edge
3:45 or another. So we can see here if it the
3:48 um appro the concept is approved it goes
3:51 to a save approved concept report and if
3:53 it's rejected it drafts for us a nice
3:55 email um to send to the designers
3:56 letting them know like hey we're not
3:58 going to go forward with this. So,
4:00 switching back to our clothing designer,
4:02 we can see here that we've gotten to the
4:05 human in the loop step and it is uh
4:07 wanting us to it says actually, hey, the
4:09 decision we want to reject it misaligns
4:12 with our brand
4:15 and um it it heavy winter collection
4:16 poorly targets audience needs and
4:19 seasonal demand. Um, a lot of risk. So,
4:21 we're going to go ahead and reject this
4:23 concept. And then what it's going to do
4:25 for us is actually create, as I
4:27 mentioned earlier, a nice email saying,
4:28 "Hey, concept designer, thank you for
4:29 submitting your clothing concept to
4:31 Zava. We appreciate your time, but we're
4:32 going to move forward with other
4:34 projects this time." That being said, we
4:36 are super excited for you guys to be
4:37 able to try out the Microsoft agent
4:39 framework and excited to see what you
