Danish Naseer: Welcome back. I hope you all enjoyed the activity we had. It was a fun time. And now we have a panel discussion with the OpenClaw team and the Cloudways team. So the topic is OpenClaw in the real world: what developers and agencies are actually building. For this panel discussion I have two respected panelists. First, Graham McBain. I would like to invite him first. Graham is a DevRel and community manager at OpenClaw where he works closely with developers and the broader community building around AI agents. Graham, welcome to Prepathon.

Graham McBain: Hey, thanks for having me, glad to be here.

Danish Naseer: Thank you so much for giving us the time. The next panelist we have is Ayaz Khan. Ayaz is part of the Cloudways product team, working on products and experiences that help customers build and run modern applications. So we have Ayaz Ahmed Khan from the Cloudways team. Hello Ayaz, how are you?

Ayaz Ahmed Khan: Hi. Thank you Danish, and thank you Graham for joining us. Really excited to have this discussion going. I’m doing fine, thank you. How are you?

Danish Naseer: Thank you so much both of you for taking out time. I think we are right on time. We can start the panel discussion. We have a developer audience and a mix of agency and e-commerce audience as well. And in the era of AI, everyone is using AI agents. For the audience, it is your time to ask questions to the OpenClaw team because we are discussing specifically OpenClaw in the real world. So we can start the panel discussion without any delay. Starting with the first question, Graham, there’s a lot of excitement around AI agents. Cloudways also launched an AI agent, by the way, for the viewers. What are developers actually building with OpenClaw today? There’s a wide range from personal agents to enterprise. So if we can start with this.

Graham McBain: Yeah, like you mentioned, it could not be a broader application of tools like this. I have a lot of friends personally, we have a lot of people in the community that are doing a use case which I’m calling the home manager, which is someone who, there’s a lot of mental overhead when you’ve got a family and a house and all of that, and so they’re just giving all of that work over to Claude, reading all the things that come in from the school, interfacing with people who are doing work on your house. That really offloads a lot of the mental load of what it takes to run a family in the world. I think that’s a really cool personal use case.

On an enterprise level, we’re seeing a ton of people use this in all kinds of applications. So there’s a guy on the East Coast who I just spoke to who has a print shop. He makes printed materials, all kinds of stuff, and he’s actually having every single one of his team members sit with OpenClaw, figure out what processes they do over and over again and how they can hand it over to OpenClaw, and he’s been able to see incredible productivity gains. We had another person who couldn’t be more opposite of what you think of from technology, but they’re a rodeo photography business. They go around the country taking photographs at rodeos, and they’ve optimized almost the entirety of their back-end software using OpenClaw. And so they can just focus on talking to customers, taking photos of people riding horses, and then all of the stuff that just needs to happen in the background just happens, which is really incredible.

Danish Naseer: Awesome. Ayaz, my question to you: from the infrastructure and product side, what kind of workloads are you seeing developers and agencies wanting to run with the AI agents?

Ayaz Ahmed Khan: Yeah, so this is an interesting space, and OpenClaw has been around for a while, but I think with agencies and specifically enterprises as well, this is just starting to take off, and there’s a lot that will be done as we go along. So a number of things that we are personally seeing with our customers, the kind of workloads, they generally fall down into a couple of different buckets. One, as Graham mentioned, is the personal assistant kind of bucket where people are using it to look at their inbox, manage their calendars, maybe they have something they want constantly watched and be informed about when it changes. So those are the kind of things people start with, and this is an activity that we’re seeing is spiky. It could generate a lot of traffic or activity at a specific period in time, for example in the morning when they want to look at their calendars and inboxes, or there’s a specific period in time where they want to see something that they periodically look at.

And then there’s the second bucket that’s more interesting from an agency perspective, and this is the thing where agencies are using AI agents to do things that are not so interesting and boring to do, but equally important for the agency to do. So something like client reporting. Nobody wants to manually do it, but you use an AI agent to completely automate it. Things like looking at your Google Analytics, your search console, other stuff, generating monthly reports, maybe creating a content pipeline. And then one of the things that we’ve seen agencies really get successful with is an SEO audit agent, an SEO going in and doing an audit of their sites and giving them a daily, weekly, specific-cadence-based report.

The third one that I’m also interested in is what we call the background kind of things where people are trying to figure out how to automate operations. So for example, maybe review logs on a periodic basis to see if there is something coming up. And then also, for example, with Cloudways managed agents, we recently added an MCP connection. So customers or agencies who are already managing their client servers on Cloudways want their agents to have access to those, and MCP is a way for them so that they don’t have to manually log into the platform and look at what sites are running, what versions are running, and so they use MCP to do that. And so that’s the common workloads that we’re consistently seeing that customers are running through.

Danish Naseer: And my opinion is, as AI agents are starting to do more than just answering questions, they are writing code, running tasks, working with tools, and in some cases operating almost like another member of the team. So I agree with you on this point. Graham, my question to you: we hear about coding agents constantly. What does OpenClaw go beyond writing code?

Graham McBain: Yeah, I mean OpenClaw, I like to think of it as the level above your coding agent. So myself and a bunch of people on our team are actually using OpenClaw to run a bunch of coding agents. So OpenClaw will go kick off a Codex session, it’ll go kick off a Claude session, it’ll go kick off whatever that is. And so OpenClaw is more like that higher level abstraction. But even beyond coding, editing, coding agents, I’ve been using it for video editing, where I’ll just record a video with one of the maintainers on the open source OpenClaw project to talk about a new feature they’ve come up with, in Riverside. And so I get really high quality recording, and then I’ll hand that raw video to OpenClaw, and there’s a skill out there called movie star which somebody created, and I’ll just say, here’s what I want this video to be about, pull out the transcript, edit that transcript, and then give me a couple different cut ideas of what you think the video should be edited to.

And then it’ll come back and say, all right, I’ve watched the video, here’s a couple different lengths, here’s a couple different angles you can use. And then it’ll say, okay that’s great, and then I’ll give it some ideas for graphics, some ideas for B-roll, and it’ll go pull that stuff and it’ll edit a full video for me. And so something that used to take me an entire day is now under an hour, and I don’t ever touch editing software. So now my video editing is just my taste. I don’t even have to worry about the tools anymore. It’s no longer this huge learning curve of learning how to use proper editing software. It’s just what actually needs to get done, and then it’s just my taste. So it’s like every middle manager’s dream, right, is to just yell at somebody until the thing comes out the way they want it to be.

Danish Naseer: Awesome. And what is the one OpenClaw use case you have seen that made you think, this is exactly what agents should be doing, other than the video editing?

Graham McBain: Yeah, one thing that we really love, we’ve got a new feature in OpenClaw called multiplayer mode, where instead of each person having their own agent on their own computer, we have one agent that’s hosted in the cloud, it’s available via web, and not only can I jump into anybody else’s session. So the other day Peter was working on a thing for OpenClaw and he built the feature but it looked terrible, and so he just @-mentioned our designer and said, can you come in here and fix this design. And so then the designer took over that agent session, fixed the design of it, and then it was really non-performant, and so he @-mentioned somebody else and said, hey, can you make this perform well, that person jumped into that agent session and made that feature performant, and then they shipped it. And so three people collaborated on the product without ever pushing it to GitHub, without people pulling it down and dealing with work trees and branches and whatever. Three people got to collaborate on a single feature before it ever left the server, which I think was really incredible.

The other thing that you can do in this multiplayer mode which we really love is you can pin sessions to the top of the environment, and those sessions, in addition to just having the conversation, can generate HTML apps and dashboards. And so we have apps for things like analytics reporting. I generated a content calendar that works, and that content calendar pulls all of my meetings, it pulls all the conversations from Slack and Discord and all the things that people are talking about about the kind of content they want to produce, and then it puts it into a real-time calendar, and I can even from that dashboard say, all right, kick off this workflow. And so all these internal applications that you might not otherwise build are just available to use and edit inside of this team environment, which I think is really incredible. And I don’t think that team agents have really taken off quite yet, but it really is the next level for this.

Danish Naseer: Awesome. Ayaz, my question is to you: from what you are seeing, where do AI agents provide genuine leverage for an agency or development team, rather than simply adding another layer of complexity?

Ayaz Ahmed Khan: So really you can use AI agents to do a lot of creative stuff, and use them to figure out new ideas and brainstorm stuff. But when you talk about leverage, particularly for agencies right now, at the stage we’re in, that comes from work that’s repetitive, that’s converted to what AI agents can do well. And so an example that I shared before is just client reporting. The real leverage there is when you do it the first time with an AI agent, it might take you a number of hours to configure and get right. But then by the time you get to your 10th and 11th and 12th client, this is only 10 minutes, 5 minutes, that’s it. Or the agent takes care of it completely. Now that’s real leverage for an agency, because that is time saved and a lot of business that they can do during that time.

And when you look at the more advanced use cases, or the more futuristic ones, you talk about radical ideas like a billion-dollar solo entrepreneur that’s building a business, while we’re not seeing that with our customers, what we’re seeing is that we have customers who’ve just built a whole agency as a single person and they’re using their agents to take care of everything. And then we have customers who are really pushing the boundary by creating a whole fleet of agents. So you have an orchestrator agent, an SEO agent, a development agent, and then they’ve connected them to their kanban boards, their linear systems, where their engineers are putting up workloads, and then the agents build them up, and then they orchestrate using those agents, and then they deliver it, and then they post updates or bring in a human for reviews.

So all of these things are possible, and we’re starting to now see agencies play around with those kinds of things, which is why I think we’re still at the infancy stage with agencies. There’s a lot that they can do and they’re starting to open up and trust these systems to be able to build it. And that’s what we offer today, is make them comfortable with this technology and have something safe and protected available where they can experiment with these things without having to worry about any of the other issues that come with it.

Danish Naseer: Making sense. And Graham, my question is to you: what are developers getting wrong when they first move from experimenting with agents to actually relying on them?

Graham McBain: Yeah, I think people spend too much time micromanaging their agents and trying to give them too much instruction. There was, maybe six months, a year ago, this idea that you had to put together a planning doc and then give that to the agent. I think you will be surprised at how much they can get done if you just tell them the outcome and not how. Oftentimes their how is going to be better than yours. And especially in an enterprise context, the intelligence is getting so cheap now that the best thing you could do is to just start multiple sessions. I think that’s actually probably the number one thing people get wrong, is they do one session, they say something to an agent, they wait for it to finish, and it’s like, no, kick something off, go kick another one off, go kick another one. Just keep handing it your to-dos, and then come back and check on each session individually, because you can get so much more done than you may realize.

Another interesting thing that people don’t do enough of is set recurring work. So just say, check on this every 20 minutes, or check on this every hour. They think of it too much as first-person. You really need to think of agents as team members, just like Ayaz said. It’s not something you need to babysit. It’s something that really can go out and do real work.

Danish Naseer: And Ayaz, what’s your take on the same question?

Ayaz Ahmed Khan: So I absolutely agree with Graham. The models that we’re at today, they’re very very advanced, and the harnesses like OpenClaw, they’re so much better now than they were before. And so if you’re just running one session using it as a chatbot, it’s a waste of this tool. This is not what it should be used for. What it should enable you to do is, again, define what the outcome or the end goal is, and then let the agent figure it out, just as Graham said. And I 100% agree with that. That’s the way to use it. Have multiple agents. Let the agent figure out what it needs to do. And it should know where it ends and what the outcome should be, and then let it figure this out. They’re too intelligent, too advanced, to not be able to figure this out. And if you are babysitting them, that’s just a waste of your time, and you shouldn’t be doing that. So definitely yes.

Danish Naseer: And Ayaz, what are the infrastructure and security considerations that developers need to think about before giving an agent broad access, because these agents can interact with tools, APIs, files, and external systems?

Ayaz Ahmed Khan: Yeah, so a couple of things that we have looked at. Of course you have to worry about your private data and content that’s not trusted, because these agents are the kind that can access stuff and then they can go out into the world. So you have to design your system in a way that enables all of this while protecting what’s important for you. So some of the things that I personally recommend is that don’t run it on your laptop where all your files are, where all your emails are, or your saved credentials are. Give it a different, separate machine, and only give it access to what you need for your job to get done. And then for things that are dangerous for you or are irreversible, you’ve got to have a human in the loop in there, whether this is through adding approval gates or something. You have to figure this out and you add it, because if you give it unfettered access then it can, for example, email one of your clients with stuff that it shouldn’t be sending out, that can be destructive to your business.

And then, because skills have become so common and ubiquitous, you have to also be careful about what kind of skills you’re putting in. So on the marketplace it’s very easy to sneak in malicious stuff because it’s all English, and then one bad skill can end up exposing the stuff that you’re building. So you have to be careful about it. And then from the infrastructure side, one of the things we’re seeing is that people leave gateways open on the internet and there’s no authentication there, there are no firewalls there, and so your API keys, your accesses, your other tokens can be easily pwned off, or they can be taken over and actually made advantage of. And so one of the things that we do at Cloudways is make sure you have a separate instance. It has an SSL available out of the box. It has a firewall there. We have other things that make sure that all of these security issues that people are seeing in the wild, while they run it themselves or they host it themselves, they do not face. And so those are the common issues that I think people who want to run agents should not be worried about, so they should just be focused on getting their job done, figuring out how to use the agents to do it, and let whatever provider that they want to work with handle all of this seamlessly so they don’t have to worry about it.

Danish Naseer: Awesome. Graham, what’s your opinion on the security part, because as we know agents and such tools work with APIs, files, external systems. So what’s your take on it?

Graham McBain: Well, you say on the security part, do you just mean how what should somebody’s security posture be for when they’re setting up their agents? Can you be a little more specific there?

Danish Naseer: Yeah, like managing their workload with those agents, like giving them the APIs and the external system access. So is it secure to give access to the AI agents, and then they can work with the agents by providing all the information to those agents?

Graham McBain: Yeah, I mean I think there’s a bit of security by obscurity here where, sure, someone can interface with your agent. I think the biggest thing is, how sensitive is the data that you’re using? Are you giving it direct access to a bank account, to internal systems? I think a lot of times, especially in an enterprise context, the best solution is to just, again, it’s funny, it’s like the answer is always treat an agent like an employee. So don’t let individual employees make IT decisions for their agents. Have it be an organization-wide discussion and design. And so letting an employee install an agent that has pseudo access to its entire work computer, probably a terrible idea. But provisioning virtual boxes for your employees’ agents that are all standardized to the same levels, same systems levels, and creating identities for those agents, and then letting your IT team manage that access and permissions for those sandbox environments, that’s probably the right posture for most organizations.

And if we think of it that way, then a lot of the tools that we have to manage security already exist. You’re not having to reinvent something from scratch. So I think it’s just, you wouldn’t let two employees use the same computer, so don’t do that with an agent. Think of an agent, just treat them like an employee from a security perspective, and then you’ll find that a lot of your IT and security policies still apply.

Danish Naseer: Awesome. Ayaz, my follow-up question to you related to agencies: for an agency managing multiple clients, what are some practical ways agents can move from an internal experiment to something that creates measurable client value?

Ayaz Ahmed Khan: Yeah. Hold on, let me pass this to Graham, actually, I think this is more of a Graham question on client value.

Graham McBain: Yeah, I think a lot of people are making a mistake here when they’re an agency and they’re trying to get customers, is they’re trying to think of individual problems that humans have and have an agent solve that individual problem. You need to, I’m going to sound like such a broken record here, but you’re going to need to think of these agents like people. And so what is a whole job that a person does, and how can an agent make this better? So there’s Sierra, a really powerful company that provides agents for people. They were doing customer service, that was their first wedge into the market, but now they’re moving on to full enterprise value workflows, and they’re doing things like loan management. So if somebody wants to refinance a loan, that entire process is getting handed over to an agent. And so it’s not just filling out forms or any low-level individual thing.

Because you’re going to have a problem where humans who have been doing a workflow for a long time, they’re very good at it, and they’ve got a great personable rapport with their customers, and that’s great, and there’s going to be an opportunity for those people long into the future. But then there’s going to be a subset of new customers that could be easily managed end to end from an agent. And so give that a try. Don’t make an individual human more productive. Take some sort of human workflow and give it entirely over to an agent. And it’s just like manufacturing. If you go from an artisan who is making something end to end and you try to standardize that into a manufacturing process, those first few runs of that manufacturing line are going to be terrible, or they’re going to take a lot longer, they’re going to be more expensive, and the end product isn’t going to be as good. But every time you refine each individual step on that manufacturing line, you’re going to start getting scale out of it that you wouldn’t get if you just hired more and more artisans. And so that’s the framing that I would think of when you’re trying to implement these types of tools.

Danish Naseer: Awesome. My question to both of the panelists, starting with Ayaz: what one agent workflow do you think will become completely normal for developers and for agencies over the next 12 months, or you can say over the year?

Ayaz Ahmed Khan: So while we’re still starting out with this, and some of the more tech-savvy agencies are experimenting with it, having multi-agent fleets is something that more agencies, once they build enough trust with agents and how to manage them, will start adapting to. And the idea is you have multiple things in your pipeline that you need done. Most of these are repetitive things that scale across different clients. They look pretty much the same. They might have some cosmetic changes, but ultimately the work is the same. So these multi-agents are going to do most of that stuff, and then you have these human gatekeepers who are going to approve things and direct them where they want to go. I think that’s the one workflow that’s going to become more popular with these agencies.

But then they have to be particularly careful about the fact that when you’re developing these kinds of multi-agents, you want those agents to have the full context about the client that they’re working with. So maybe they should understand the tone of voice that that specific client speaks, or the kind of site they have, the kind of analytics they want to do, and other specific SLA requirements that they have. But then they also have to make sure that these are protected well enough so that there are no breaches across those clients, so one agent goes in and does something to another client. And so that’s something that they have to get more comfortable with, and look at what are the different workflows that they can adapt to prevent this. And so particularly at Cloudways, that’s one of the things that we initially started to do, was having instances that are dedicated to specific clients. I do think this is something that agencies are going to start with, and eventually have a lot of visibility into both pricing and performance and other aspects of hosting and running an agent, and eventually from there they’ll shift to a multi-agent approach where they can get that real leverage.

And as Graham was talking about, even in the enterprises that we’re looking at, they’re really struggling with that transformation part, because the workflows are really messy. When somebody is going in and trying to look at how do we convert that workflow into a number of agents that are doing it, you really need to understand how those workflows work, understand each and every step that they take, and then convert them over to agents. And so you eliminate all those low-hanging fruits, and one by one you start transforming all of those workflows that can be. So that’s an approach that I do think is going to become more common in the next one to two years. But yeah, multi-agent is where the agencies are going to get the most leverage out of.

Danish Naseer: Awesome. Graham, what’s your take on the same question, the workflow you think will become completely normal for developers and agencies?

Graham McBain: Yeah, I think from an agency perspective it’s going to be account management. I think that you can focus entirely on getting the work to be done well, and I think you could do a much better job giving account management over to an agent than to an individual. It’s going to be more responsive, it’s going to have better context, and most of the time you’re not going to be interfacing on a video call, and those few and far-between video calls can be handed over to the principal of the agency, which I think makes a ton of sense. From developers, I think we’re probably going to solve code review. That’s the biggest blocker right now in AI coding. And so code review is going to be the place where we’ll start to see, okay, not only is it passing all CI tests, but it’s also auditing itself down. Right now people are seeing that Astra is being a little too verbose. We found at OpenClaw that one single line of code had three tests on it. It’s like, just a little excessive. And that ups your CI bills like crazy. So I think that’s going to get solved for sure.

Danish Naseer: Awesome. Graham, if someone watching this has never built an AI agent before, what is the first real-world problem you would tell them to automate?

Graham McBain: That’s a great question. I think what I would do if you’re building an AI agent from scratch, I would pick something that you do every day. So first, before you actually build it, I would use an off-the-shelf AI agent, and give it your schedule and process and say, what am I doing here, what do I need to build to fix this? And that’s going to give you a pretty good idea. I think humans are really bad at self-reflection and discovering what they can do better. And so ask for third-party input from an agent of what you can do better, and then go through the process to have an agent build it for you.

Danish Naseer: Awesome. Ayaz, I have one question for you, then we can wrap up this panel discussion, that is related to building a demo. Building a demo is one thing, running an agent continuously is another. What changes when an AI agent becomes a production workload? If we can discuss it, then we can take some questions and wrap up this panel discussion.

Ayaz Ahmed Khan: Yes, so a number of things come up. A demo, and that’s also what’s happening with enterprises, that they mostly don’t go beyond those demos. But in particular for this specific case, for example, memory and CPU resources. A demo does not sufficiently demonstrate what happens when an agent is left running for a while, how the memory builds up and how resources come up. And if you have, let’s say, a 2GB server running OpenClaw, and that memory and resource usage goes up, it can actually crash it down. So a demo will not show you how the resource usage comes up. And then the other thing is token cost. In production, depending on what you’re trying to do, those can really go up. And depending on what models you’re using, if you have more state-of-the-art models configured for something that does not require state-of-the-art models, it is going to eat up your budget, and that would be a very shocking surprise for you.

And then updates in particular. I know OpenClaw operates at a pretty fast iterative frequency, and more software is going to follow that up and it’s going to get faster now. And so how do you keep up with changes that break, or incompatible changes in skills or configs and stuff like this? A demo obviously does not show you, once you get into production, you have something set up, a new version comes in, you run into those issues. So you have to figure out how to do this. And that’s also one thing with Cloudways, is we vet new releases to figure out if there’s any regression and how do we proceed from there, and if there are cases where something’s going to get corrupted, we’re going to figure out how to do this. So that’s something to take care of.

And then the other thing is, when you have agents running and they run into issues, they’re not going to crash and cry about it loudly. They’ll just quietly go away. So your workflow is going to completely go quiet. And so how do you build workflows or systems around it to make sure, if they’re not performing the way you want or they’re running into issues, you get alerted that something’s happening. I think those are the four main things that really come up initially, but those are the main ones that we’ve seen really get into production that bother customers, and then they either just give up, or they spend enough time trying to figure this out. And while we designed Cloudways managed agents, this is some of the things that we try to cater to. So with automated backups and one-click recovery, firewalls, and vetted versions and stuff like this, to make it as painless as possible for customers so they don’t have to worry about the systems administration and the DevOps kind of thing, and just focus on continuing to work on their workloads and figure out how to use more of agents to get that leverage out.

Danish Naseer: Awesome. I was checking the Q&A session and there are a lot of questions for the OpenClaw team, I think, but we are right on time. We asked the questions to Graham and Ayaz. So for the audience, if you never tried OpenClaw, I think you can try it out. Cloudways is also providing the AI agents. We already placed the banners on the booth, so you can get there and check out the AI agents Cloudways is providing. We are also providing hosting credit to get started free. I would like to thank Graham and Ayaz for the wonderful panel discussion around the OpenClaw real-world scenarios, and thanks again to both of you for taking out time for this panel discussion. We are right on time. We have another activity and another session lined up. So Graham, thanks to you for taking out time, and Ayaz. And I would request you both to stay active in the chat if you can, because there are a lot of questions for both of you. So it would be great if you answer all of those questions, or a few of those. So Prepathon audience, that’s Graham McBain from the OpenClaw team and Ayaz from the Cloudways engineering team. We are done with another panel discussion with the OpenClaw team. I think we can move towards the next session, that is with Tracy Lee. We will be back in a short time. So stay tuned. We will be back with a session with Tracy Lee.