Why Managed Agents Matter

SPEAKER_00 0:22

Today we're going to be talking about building managed agents in Claude. This is a pretty cool topic, so stick around. Once again, welcome back to the Digital Customer Experience podcast with me, Alex Turkovich. So glad you could join me here today and every week as we explore how digital CS and AI can help enhance the customer and employee experience. My goal is to share the insights that my guests and I have learned over the years so that you get what you need to evolve your own digital program. If you'd like more resources, head over to digitalcustomer success.com. There you'll find access to a chatbot that's trained on this content and get information about the Digital CX masterclass where I teach you how to execute and thrive in Digital CX. For now, let's get started.

Platform Notes And Tool Parity

SPEAKER_00 1:08

Hey, Edward Alex jumping in here before we start this episode. I wanted to jump in real quick and talk about the fact that what I'm doing today isn't Claude. In fact, I think all of it is Claude-based and using managed agents from Anthropic. That said, most platforms out there these days have this sort of feature where you can build agents. I know you can do that with OpenAI. I know you can do that with Perplexity Computer, for instance. A lot of different platforms now have this kind of agent-building capability. But what I lay out in this episode holds true no matter what platform you're using, because fundamentally agents work in a very similar fashion across the board. They all need context, they all need an environment to operate in, they all need to understand when they should be kicking off, what their core instructions are, and those kinds of things. That doesn't change. The nomenclature changes slightly depending on the platform you're using. But fundamentally speaking, it's all the same stuff. And so this episode is applicable to you, whether you're a clawed user like I am, or whether you're using codecs or computer. So with that said, let's dive

From Scheduled Tasks To Agents

SPEAKER_00 2:25

right in. Hey, my name is Alex Turkovich. Welcome back to the Digital CX podcast. This is episode 110, and we have a pretty cool episode in store for you today. I'm gonna probably spend a few future episodes, including this one, where we're kind of kicking off this project of standing up Claude Managed Agents. Now, if you're a regular listener to this show, you'll know that I don't know, several months ago I talked a lot about scheduled tasks in Claude. So if you are a heavy Claude user, scheduled tasks was a great way to start getting into agenc work and working with Claude in a way that was agenc because these scheduled tasks would, you know, kick off on a regular basis and do work on your behalf without you having to basically kick off that workflow. That is, by in essence, the the definition of what agentic AI really is for the most part. There's some more detail there, but we won't get into that. I have recently been playing a lot with Claude Managed Agents, which was released a few months ago and takes this to a whole nother level because what we're able to do with managed agents, which is really more on the back end of Claude, is pretty phenomenal because you start thinking, you know, everybody's kind of talking about, okay, I've got this team of AI agents and I've got this group of AI agents working together and all that kind of stuff. And so what I figured I'd do is, as I'm building this up for my own business, is kind of take you along for the ride. And so today we're going to start at the ground level and talk a little bit about what managed agents is. And then we're going to dig into a little bit more about the use cases and how I'm planning on using this. And of course, I would love to solicit your feedback on are you using managed agents today? Are you not? If you are, what are you having them do? How are you having them talk to each other and all that kind of stuff? So before we get more into this, let's talk a little bit about what the heck I'm talking about.

Where Managed Agents Live

SPEAKER_00 4:34

Now, if you are listening to this instead of watching it on YouTube, for example, I'm going to try to be as super descriptive as possible so that you can kind of follow along. But what we're talking about here is not in cloud.com or your cloud desktop app. These are essentially API-based agents that you can set up on the back end of Cloud. And specifically, if you haven't even been there, platform.cloud.com is a pretty cool place to poke around because it is where a lot of the backend stuff that you work on every day kind of lives. You know, if you are a regular cloud user and you you haven't dug into platform at all, I would highly recommend you take a weekend and go do that. Um, but today we're going to talk specifically about managed agents in platform. You get there by going to platform.cloud.com. You use your login and then go to the managed agents tab. Now, one thing that is pretty important when it comes to starting to work with this stuff is let's say, for example, that you are a pro or a max cloud subscriber, the thing that you want to avoid is just going nuts on it because we're not actually using usage towards those subscriptions. We're using API based token usage. And so you do have to fund the, you know, you do have to fund your account so that you can start using tokens. And so that's where once we start building these agents

Token Costs And Model Choices

SPEAKER_00 6:09

and they start nesting under each other, that's where selecting the model type is going to be very, very important so that you're not chugging through tokens unnecessarily. I know it would be great to have everything, you know, running on uh, you know, Opus or Fable or whatever, but that's a quick way to run up a pretty big bill. So you want to try to avoid that. Now, let's talk a little bit about what the components are of these managed agents. First off, you have the agent itself. The agent is essentially a set of instructions, right? It is a hopefully a pretty narrow task that you want this agent to accomplish on a regular basis. Let's uh give you a couple examples. One is, I don't know, reading your

Agents Environments And Memory

SPEAKER_00 6:55

email, looking through your email, highlighting things that you need to respond to, or maybe flagging certain emails that it wants you to look at. So that would be one example. You know, another small example would be well, let's take this podcast, for example. I'm going to be building an agent that once the, you know, once the audio is ready and the transcript is ready, it's going to take that transcript and turn it into a draft for the show notes, for example, as well as a draft for the YouTube description, as well as chapters, you know, a set of chapters for that. And so it's kind of like this. I'm going to have one agent that does the package of that essentially. I am going to implement a whole series of other agents that are kind of wrapped around this business. And I'll actually kind of share with you what that is in a second, but that's essentially what an agent is. It's a task. You want your agents to be somewhat focused on specific tasks. In other words, you don't want one agent to do too many different things because you can create a multitude of agents, right? And what I will be doing is creating a multitude of agents that then essentially report to an overall agent or a kind of a chief of staff agent, essentially, that then takes the outputs of all those agents. And that's the one that I will be interfacing with on a regular basis. So that's the agent. The other kind of definition to know here is environment. When you go to set up these managed agents, you're going to want to pay attention to the environment in which they run. And so you can set these environments up and they allow you to restrict what the agent running on that environment has access to. It allows you to restrict, you know, what tools it has access to and and what areas of the internet it has access to. So that's essentially, yeah, I mean, it's pretty self-explanatory. There is in Managed Agents also the concept of a credentials vault. So invariably, these agents are going to want to work with your other tools. And so if you if you need it to work with others, other tools, you can create a vault of credentials. So let's say maybe your HubSpot API key is in there, or your Stripe API key is in there. And of course, just like with any API integration, you can restrict and limit what each one of those API keys gives, eventually gives the agent access to. So you have a credentials vault. Memory stores within managed agents is the place where your agents are going to be storing their memory. And so what you will have over time, the more these agents run, is a memory of what they've done in the past. And those agents will then be able to essentially self-improve upon themselves based on what those memory stores contain. So that's super, super cool. And the last little definition basically that you need to know about is um sessions. Sessions is basically like your chat in uh chat GP, you know, chat GPT or cloud or whatever. But it's it's you know, one when a session is running, that is one session that an agent is running for you, right? So if you have 10 agents, they run, they each run once a day, that means you're gonna have 10 sessions running essentially. So that is the kind of definitions element of you know, what all the the components of these agents are. And so when you go to build these agents, which is super easy because you can just prompt it out, and I'll I'll kind of share what I've done to do that with you. When you build these agents, you're then going to also define what environment they're in and what credentials vault that they use. So that gives you then the control essentially of what they have access to and then how they integrate with those things.

Audit Your Work With A Prompt

SPEAKER_00 10:51

Now, staring at a blank screen about to create agents can be a kind of a daunting thing, right? And so uh, you know, I've been steadying up on this for quite a long time because you know, these these are the kinds of things that I like to really dig into and and get into YouTube university on. And uh a couple of things that stood out to me when I was researching this is there was a few, there were a few people who suggested that you essentially do an audit of yourself and your use of Claude and your use of AI to understand what would be a good candidate for an agent. And so I did exactly that. And I'm gonna give you, I'm gonna, I'm gonna read you out the prompt that I gave Claude. I'll also put this in the show notes so that you can go copy it if you would like. But my goal here was, you know, I'm always trying to automate what I'm doing here with the podcast and the masterclass and you know, the newsletter and all those kinds of things. While my authentic voice needs to be there, and I spend a lot of time writing and putting posts together and obviously recording these podcasts. The thing that I need efficiency on is like emails and and you know, just the day-to-day stuff, creating drafts for me and and and doing a variety of different things. And so what I wanted to do was basically have Claude analyze my usage of Claude as well as what scheduled tasks I have running currently and all the other things that I do on a regular basis to provide with me a list of suggestions for agents to create. And so here's the prompt that I used. I said, I'm wanting to automate my business via a series of claude-managed agents doing a variety of tasks that are currently being done either manually or via scheduled tasks in co-work or routines in Claude Code. I want to be quite niche with those agents so that tasks can be delegated to a high degree of detail across those agents. That means I'll also need some agents that coordinate with each other to get broader tasks done by their subteams, much like a team structure works in the workplace. I would like you to do an audit of my business based on conversations I've had with Claude and tools that you have access to. Then create a detailed list of agents you suggest be created to help automate and coordinate this work. Also, make suggestions for additional tools needed to execute this work. However, try to manage within the tools that we already have in our tech stack. Ask me any clarifying questions you have until you have 100% certainty of the task you need to accomplish and that you'll execute it to a high degree of accuracy. So that is the prompt that I will include in the show notes, but that is the prompt that I gave it. It asked me two rounds of questions, mainly around uh, you know, more finite scope of what I wanted it to, uh what I wanted these agents to be. It also asked me whether I wanted to exclude certain things that weren't really in the realm of what I do, but I just asked Claude about. Of course, I, you know, I asked Claude about a bunch of stuff. So I didn't include any of that. It asked me how much autonomy these agents should have. And I basically suggested that we do somewhat of a tiered, you know, set of autonomy behaviors based on the risk involved in some of this. Um yeah, and so, you know, it asked me then a second round of questions around what manual work that I do that it may not uh be able to see. And so some of it was like accounting related and administrative related. And it asked me about budget, so that was very important. Obviously, when you're running this kind of stuff, you don't want it to go super, super, super wild. And so I gave it kind of this parameter of budget that I wanted it to stick within. So then what it did is it essentially spit out a suggested blueprint for me and my uh agents. Now, what it did is it gave me a whole list of the things that I do on a regular basis. It gave me uh, you know, kind of some operating rules. So, like the risk tiers that we talked

The Org Chart Approach To Agents

SPEAKER_00 16:59

about, things that it would just do automatically for me versus things where it would like just draft things for me and things where it wouldn't touch anything at all. Then it started to give me essentially an org chart. And so when you think about, you know, uh a podcast like this and a show like this, there are a variety of different things that go into it. Um, there's much more behind the scenes that goes into this than just me sitting here yapping in front of a camera. And so what it did was it basically gave me the parameters of what a chief of staff would be. And so this is a chief of staff that is over a variety of different work streams and kind of dictating a certain set of things. And then it prescribed a series of uh desks, if you will. So, you know, one of them was a content studio, and so the the agents within the content studio, it it recommended a topic scout. So a scout that would go out and look at uh you know things that are being talked about in CS, happenings in CS, and then it would cross-reference it with my knowledge base that I've built around this show and everything I've talked about. It gave me, you know, it per kind of social channel, it gave me a writer. So like somebody who would uh an agent that would basically draft things for me. I don't think I'll ever get to the point, by the way, where I have an AI just post stuff for me. Like drafting, I can do, but straight up posting, I don't think so. It gave me a clip producer. So I already use a tool called Opus Clip to create clips for me of uh fully published episodes. However, this clip producer then does some analysis on those clips and helps to or makes recommendations for which ones that should be posted and then also what kind of captions should uh go along with that. Uh it recommended a an agent for thumbnail and graphics creation. It uh recommended an agent for a queue scheduler. So, you know, queuing up episodes and really planning those out. Again, it seems like a small task, but in the aggregate, when you do it over and over again, it's a it's a it's kind of a big thing. And then also like a monthly trend scout. So, what are the monthly trends that it's kind of seeing out there that we might want to talk about? And that's just the content studio. There were 36 total agents that it recommended. So that was the content studio. There's a podcast desk, which one of them that it suggested, which is quite interesting, is a pitch triage. I can't tell you how many emails I get from like podcast placement services about, you know, people that should be on the podcast. And look, I'm I'm all for folks reaching out and saying, hey, I've got this cool thing that I'm working on. Can I be on the podcast? I love that kind of stuff. I pretty much flat out ignore anybody who uses like a podcast pitch service because there's a ton of them out there, and I get probably five of them a day, and I just ignore them. And so it suggested an agent that actually looks at those pitches. And if there are guests that do make sense, okay, maybe I can engage those. But oh man, it's crazy. One of them is a guest coordinator, so an agent that actually handles some of the coordination that happens before uh recording an episode and after, show notes writer, knowledge base librarian. So, you know, there's all these kind of sub-agents that do really minute things. One of them is a growth or another desk is a growth desk. So that has, you know, agents that are working on funnel and nurture campaigns and kind of more marketing related things. Then I have a revenue desk. So that's like, you know, a sponsor, prospect, or sponsor account manager. I don't know if I'm gonna do an agent for that, but whatever. These are just Claude suggestions. An invoicing clerk, so somebody who takes care of some of the invoicing and sending them and following up on invoice invoices. Now for the master class, it actually made a couple of suggestions that I really like. One of them is student support. So if a student in master class needs support with certain things that are maybe more technical in nature, can't find certain things, I want to you know have an agent able to provide some base level of support there at all hours. Um, a testimonial collector. So if somebody goes through masterclass, I have a video platform that allows me to collect testimonials, super cool. Takes manual effort to do that. So maybe I have a testimonial collector that does that for me. Product operations. It takes a lot to maintain a course like that on the back end. So you have you know registrations and all kinds of different things that happen on the back end. And so basically having somebody to monitor all of that. And then lastly, it's it's back office stuff, it's you know, kind of bookkeeping and spend auditor and all those kinds of things. So it, you know, again, you can do this as well by using a prompt or something modified off of my prompt to have Claude do an audit of your business and your daily life and what it is that it feels like it could do a good job of with an agent. And from there, it's about creating these agents. And I've started to create a few of these that have made sense. I'm not gonna implement all of those. I think that would be a little bit overkill, but I think you can tell from what I just talked about that there are certain ones that I think you know could be quite good. I do want to include a few more that help me triage my email a little bit better, help me triage my calendars, yes, multiple calendars a little bit better, help me manage just that day-to-day a little bit. One of the things that I'm really sad about is that, you know, there isn't much out there for LinkedIn. I know there's some platforms that that do kind of work with LinkedIn. And I have, you know, I do have a LinkedIn like post automator tool that I use that is quite good and does have have an MCP, but that's really more for pushing content out, less so about engagement and communication. Now, I think that's actually kind of a good thing with LinkedIn because it's hard to be spammy in LinkedIn. It's hard to, you know, unless you're using some kind of you know computer use agent or whatever. LinkedIn makes it really, really hard to be spammy on LinkedIn. And I think that's generally a good thing because if agents could do, you know, all kinds of different marketing y things, I think it would kind of erode LinkedIn a little bit. All of that aside, that's how I'm planning on using management. Managed agents specifically for this business. Now, when it comes to a CX perspective, my mind just goes

CX Use Cases For CSM Teams

SPEAKER_00 24:01

nuts with all of the different use cases that you could use managed agents for in a CX environment. Think of customer success managers. I mean, we've we've all been talking about QBR prep and those kinds of things forever. I think a couple of the things that an agent can really help with is the monitoring of accounts, monitoring for signs of trouble and churn, monitoring for wins, things that customers are doing well. And so if you're in an environment that does allow you to integrate with, you know, Claude, I would highly recommend using some managed agents and some very specific ones per CSM so that they can manage their book of business in an agentic type of way. So what I would what I would suggest is you you create essentially a chief of staff per CSM. And that chief of staff has access to these sub-agents that monitor their book of business, monitor their renewals, monitor their upsells, monitor for signs of trouble, monitor for adoption trends, monitor for upsell opportunities. You know, I mean, we we can go nuts with this stuff, really. The point, though, is giving your CSMs some breathing room when it comes to digging through reports and giving them the lay of the land, maybe on a daily basis, where you know, they know what's happening across their book of business based on what your business's specific criteria is. Now, I know a lot of CSPs kind of do some of this. I know there's a lot of evolving technologies that kind of do this out of the box. However, they may not be connected to your entire tech stack. There may be some different things happening. There may be some environmental things where you want to specifically ask about certain things that maybe your CSP doesn't have access to or whatnot. This would be a really great way to automate that for your CSMs. As a leader, I find Claude to be incredibly helpful to get me that overarching high-level view of my entire business. So looking at it from a management perspective, what are certain CSMs doing well? What are they not doing well? You know, I manage a support org, I manage an implementation org as well. And so, you know, I'm going to be using these managed agents to manage stalled implementation projects, uh, for example, or um, you know, customers that are graduating implementations. So I can reach out and say congratulations and welcome them and do all that kind of stuff. All the stuff that's really hard to do manually and take a lot of effort to do manually will now be able to build these agents around. And sure, we've kind of had this ability for a little while. And I've I have quite a few scheduled agents set up or scheduled tasks, I should say, set up in Claude Cowork already that do some of this stuff. However, pulling it all together in a set of managed agents gives me a sense of coordination between those managed agents, not just separate tasks firing off and creating kind of more noise. This gives me, through my chief of staff agent, gives me a centralized place to look at this stuff. So,

Centralized Coordination And Next Steps

SPEAKER_00 27:19

look, I'm going to be building some of these things in the coming weeks. I'll take you along for the ride because I'm a big fan of like building in the open and sharing this kind of stuff. If you've played with managed agents in Claude, let me know. We'd love to hear what you've been building, what you've been working on, so that I can benefit from that and maybe pass along some of those things as well. I know there's a bunch of folks who've been building stuff with like HubSpot agents as well and whatnot. So if you're building agents elsewhere, let me know as well and how that's been working for you. If this is something where you are kind of mondo confused and don't know where to go and don't know what to do, I will be doing some screen shares in the coming episodes that walk you through this kind of stuff. Because I mean, I'll be honest with you, setting up a scheduled task in Claude is a relatively easy thing. When you get into the platform side of things, it becomes a little bit more technical. Now, Claude can help you through all this. It's totally doable, even as a layperson. You don't need to know any code or anything like that. But it does get a little bit more technical, especially if you're working with God, Gmail and Google and all that kind of stuff. Like the API keys get a little bit weird. So anyway, let me know your thoughts. If you're working on this stuff, I'd love to hear from you and love to share my experiences with you as well. I hope you've enjoyed this episode. I hope it's given you something to think about in terms of how to automate things a little bit more. And we'll catch you next week on the next one. Have a good day. Thank you for joining me for this episode of the Digital CX Podcast. If you like what we're doing, consider leaving us a review on your podcast platform of choice. It really helps. Don't forget, digitalcustomer success.com is where you can find extra resources, a chat bot trained on this show's content, and the Digital CX masterclass that helps you master digital and AI in CX. My name is Alex Turkovich. Thanks again for joining, and we'll see you next time.