AB Workflow Agents MVP

[00:00] Okay, exciting stuff here. So I have an MVP of the barometer workflow agents built out and I'm excited just run through this real quick with you. So what did I

[00:10] do? First thing is I had Cloud Code build out the file structure that we need. So

[00:20] this is inside of the AI shared drive. There's all of these agentic barometer ones and there's agentic barometer workflow. I didn't use the little

[00:30] separator because that's not code friendly. So Cloud Code rejected it basically. So it wanted it nice and clean. So inside of there is the agents

[00:40] codecs and ingestion folder which we could delete. We can leave for now. It doesn't really matter. The outputs folder and then starters folder and I'll sort

[00:50] of go through what's inside of each of those things in a second. So that was the first task. I sort of build out the MVP. I basically told it this is what we're trying to do. This is the the setup and everything. This is kind of what I want

[01:00] how I want it to work. It's going to be you know at this point they're basically manually run tasks. So there's no scheduling set up right now although the

[01:10] scheduling framework is built. So if we wanted to add that later we could but for now like basically the agents are aware that they could be scheduled or they have a sense of how often they're running but we're manually running them.

[01:20] And then the other pieces that build out a bunch of context. So you know the the current trends, goals and methodology, guardrails, recent topics, trusted sources

[01:30] and voice and tone. So these are all things that you know I had to basically infer from the documents and stuff that were there but are all things that we

[01:40] want to actually confirm and and tweak right. This could be a lot of our work over the next coming months could be just tweaking the guardrails and the sources and the goals and things. And then it of course has the agents. So

[01:50] there's some global roles for the agents. There's a drafting agent, a methodology agent and the research agent. This onboarding agent is actually for you

[02:00] Stephanie once you have your sort of like wits about you in this environment you can run this and it will go through everything that it made and basically ask you a series of questions and confirm all these simple things. Ask for

[02:10] your input and then when it's done we'll go and make all those edits to the all the appropriate files. So that's a big thing to remember here. All of these files are written but we don't have to be responsible for keeping track of them

[02:20] or tweaking them. They're written in a way that they refer to each other so if we need something changed or if you don't like the way that a certain part

[02:30] of the report is being written or something's happening you can simply tell cloud that and it will go and find the right places to make those edits. That's a big critical piece here. So we're working with files now right that

[02:40] Claude in the cloud is writing and processing and thinking about but ultimately live on our computer and then we are conveniently saving them in a

[02:50] Google Drive such that Google Drive completely separate of Claude is syncing them between our computer. Just a nice easy workaround for avoiding this sort

[03:00] of teams thing. So that brings me over to Claude. So in Claude I basically did two research reports. These would be the things that run on a daily basis. I ran

[03:10] two of them just to see what happens in Claude code versus what happens in Claude teams. Both of them were run with Opus 5 at medium effort and the long and

[03:20] the short of it is the code one was just a much better experience and a much better output and honestly the real difference is basically just clicking

[03:30] this tab like it's a really is basically I'm with no difference. I'll show basically what happened and why. So I ran the first one with Claude code and you

[03:40] can see this is the workflow. I you know how does this work? Sorry was this one you basically go over to starters and you open up the research agent starter

[03:50] MD and you can see it has some instructions in here and I basically copy this pink piece and paste it in and so that's that pasted that's what it

[04:00] looks like and I hit run. Now it went ahead and ran the command it ran it read the files and it used a tool so you can see it's finding all these things. Use the tool is really critical here it has the ability to use and write its own

[04:10] tools which is what teams doesn't have. Then it did its research cycle browse the web research complete wrote in the report. In between it asks me a couple

[04:20] times like can I go and check this website can I check this website blah blah blah and after a couple times I was like enough go do your thing and so I can in code which you can't do in teams selected the mode and put it on auto and

[04:30] auto basically go ahead and says if this is a safe action I'm gonna head and do it and so I can just let it run and then when it's done it came with these

[04:40] three findings that matter the most out of its research report showed some kind of honest limits things here and then the three you know wrote out the files

[04:50] or whatever I can then look at that and basically it wrote this report all right so there's its whole research report that it it ran

[05:00] This is a screen recording transcript for a Voom video transcript for a Voom video

[05:30] transcript for a Voom video transcript for a Voom video

[05:50] come back and hit, you know, approve, approve, approve. And

[06:00] so that wasn't as great. And then it also failed at the web scraping for a handful of them. So if it can't do a search, right, like cloud code will just run its own scrape or do

[06:10] something and then get the information it needs. Where the cloud teams environment is basically like I can't get this, it got some of this the information when you look at the report, but you know, I guess it's aware that it wasn't as

[06:20] comprehensive as it could have been basically. So that's the that's the sort of take. Those all the reports are written into

[06:30] the outputs folder. So you'll see an outputs research agent. Here are those two MD files, the one that I just pulled up there

[06:40] and then V2, which is the one that teams did. And so you'll see like if I look at them side by side, like they look, you know, pretty much identical. There's some formatting

[06:50] differences like code, cloud codes followed the formatting better when it comes to actually building the pieces and stuff. But again, you know, it's like all of the unless we're really

[07:00] reading it and paying attention to it. As long as all the token information is there, they'll get ingested either way. So it's not that big of a deal to like make sure the formatting is perfect for these reports at least. So yeah, so then the next

[07:10] step would basically be to set this up. So if we did want it to be scheduled, I wonder I'll take a look at how this works as far

[07:20] as the auto approved thing. Because a difference between codecs in here is Oh, maybe I can do it as a routine. Yeah,

[07:30] okay. So this could be scheduled either way, we can deal with that later. But the next piece would effectively be the agents or the next agent, which would be the methodology agent. And

[07:40] this is the one that would actually go and start to look through all of the stuff that's there. Right and read all the trends, read the sources, and then go through and find those

[07:50] those trends and things that matter for the report. So again, doing that onboarding piece will help improve the methodology agent dramatically. Like I think that is probably the most IP

[08:00] heavy is the methodology agent here. If I were to already guess so. And then like I said, this was these are both done

[08:10] with Opus five at medium level. If I look at my usage, I've burned about half of my allocated tokens per five hours.

[08:20] And I'm on promo right now. But still, it's not not a crazy amount, like I could have gone through all of them. The setup was included in this and I burned about 150,000 tokens

[08:30] doing the setup. So, you know, I would say that each run is probably not going to be significant enough to really matter at the moment. But yeah, so that's that when it comes to

[08:40] setting it up, because definitely, I think that's the next piece, I would encourage you to skip the teams thing for a moment and try the code environment. I think you'll like

[08:50] the experience a bit better. The way that you would actually set this up. Oh, looks like I haven't pasted here. Okay, perfect. So you'd come over to new. And you can see I already

[09:00] have this recent one selected. But all you would come down here to do is hit open folder. And you can see it pops open my

[09:10] Google Drive. So I go into share drives AI, and then click this and hit open. That will basically add it as a kind of project it by default over here, it will be grouped by date. So

[09:20] they're all going to look like threads. But if you go in and you group by project, it'll align them based on the project

[09:30] folder. And that can make it a little bit easier to say like, I'm always working with this set, you know, of files and stuff. And basically, once you have that, then you can paste in

[09:40] the prompt. And I would encourage you to move it from accept edits to auto. And then go ahead and hit hit go and see what happens. And you'll notice it'll go through and do all its

[09:50] things. And then it'll start actually writing out the file that it needs to. So that's that exciting. I'm curious to see

[10:00] Your thoughts here, how it works, all that fun stuff, and look at that, already 223,000 tokens. Not bad. Okay, talk soon.