Building Multi-Player AI Systems with MeshAgent // Tula Masterman // Agents in Production 2025

Adam Becker [00:00:00]: [Music]

Tula Masterman [00:00:08]: Thanks so much, Adam. And thank you everybody. I'm really excited about the opportunity to be here and to get to share a little bit more with you about MeshAgent. So first, just to introduce myself, my name is Tula Masterman. I'm an AI agent solutions architect here at Mesh Agent. In previous roles, my background was more in Data Science. As ChatGPT took off and generative AI has really dominated the AI industry and really every industry over the last few years, my focus shifted to building agents and doing a lot of research into other agent frameworks, trying to understand, you know, what are the best practices? How do you build a good single agent system? How do you build a good multi agent system? How do you actually get this, you know, deployable in somewhere where people can really get their hands on it and use it in a way that's useful and productive. And I recently joined the Mesh Agent team and I think one of the unique things about MeshAgent is our focus on really enabling your entire team team to interact with agents instead of just focusing on that one person and their relationship with an agent.

Tula Masterman [00:01:10]: I think today, most of the time you see that an agent framework stops at that one person. You might have one person who's talking to one agent and that agent's perhaps really good at executing a specific task, and that's great. And you might also see cases where you have one person who's talking to a team of agents. Right. There's a lot of different frameworks out there where you have an orchestrator agent and then you have different agents that are, you know, focused on other tasks kind of within the overall agent chain. But I think the problem that a lot of people aren't addressing yet is, well, what happens when my entire team shows up and I have multiple people who need to be interacting with these agents and actually all building something together collaboratively. And that's really where Mesh Agent comes in. And that's ultimately the problem that we're focused on so solving.

Tula Masterman [00:02:02]: So Mesh Agent is your end to end platform where you can both build, deploy, share and operate agents collaboratively. And our goal is that you can build agents in any way that you want to. And we'll handle all of the infrastructure and scaling behind the scenes to actually make it possible for your teams to interact with the agents in the way that they need to. Excuse me. And we do all of this by uniting your teams, be that both humans and agents, as well as tools, as well as different data sources within this concept of a secure room. And so you can Think about a room as really the security boundary, and all of the context that's within a room will stay there. And this room is where your team can get together to actually do the work that they need to do for any given project. We realize that most of the time, any business process is going to span multiple different roles and types of people, and it's also going to span a variety of different systems.

Tula Masterman [00:03:05]: And you need humans and agents who are able to interact with both of those systems and share information easily with one another in order to achieve your goal. And so that's, that's really the, the key focus of the room is kind of providing that connection point for your team. And. Sorry, so, so within the room, one of the unique things is that the communication infrastructure is built to work for both people as well as for the agents. So we wanted to make sure that you can message, you know, other teammates as well as other agent teammates the same way, because you can kind of fall into this trouble where you maybe have one set of infrastructure that works really well for the agents, and then you have a separate infrastructure component that works well for your human team. But then the moment that you try to bridge them together, all of a sudden things don't work anymore. And so in the room context, we really do treat agents the same way that you would treat a human teammate. So that way you can have kind of the same operations work across all participant types.

Tula Masterman [00:04:14]: One of the other great features about the room is that it will start the moment that any participant joins, whether that's a human participant or an agent participant, and then the room will spin down as soon as the last participant leaves. So you're not worried about infrastructure running when you don't need it to be. And you can simply save any of the documents or artifacts from the room to, you know, other external storage. If you need, you know, maybe a document needs to get written somewhere to its final home, you can do that as well. So that way the information that needs to exit the room and be saved in your systems can be. And just to kind of ground this in a real world example, to show, you know, what this might look like from a team setting, I wanted to take the example of a health insurance company responding to an rfp. So if you have a health insurance company, you know, every year they're getting a ton of RFP requests from different folks who are looking to figure out, well, you know, what's the group healthcare plan that we can provide to our employees? And there are a ton of different roles that actually need to go into responding to that RFP to make sure that they have all the information correct and they're able to price things correctly and that they're legally compliant and all of those different components. And so you can see that in this example, having a secure area where all of these people can seamlessly collaborate together would be a big help because you might have different agents that the different folks on these teams need to interact with.

Tula Masterman [00:05:46]: But ultimately, this all kind of boils down to having one shared deliverable, which is that rfp. So within the room, you could invite people from the underwriting team, from the actuarial team, from sales, legal, etc. And all those different roles can be represented in the room. And then you'd also have the agents that help them perform the different functions. So maybe like the actuary team, they might have some agents that they're working with who are helping them to run, sort of run some sort of like risk or liability analysis. And that might involve something that uses like a custom AI model. So you might have an agent with a tool to call a different custom model, and then that response gets returned and saved in the room, and then can be put into the RFP with document writer agents or other things of that nature. And the nice thing about this is, as long as there are participants who are working in the room, the RFP can start coming together and your human team can go in and out of the room as they need to to check progress.

Tula Masterman [00:06:48]: Or if the agents need help, they can always send some sort of a notice to the person on maybe the underwriting team. Excuse me, or on the sales team, and they can request them to, hey, we need you to come back and provide more information. So it makes it a very seam, seamless way for the entire team, both human and agent, to collaborate together inside of this room. And then similarly, let's say, you know, you have all of these different agents that are designed to help with a variety of functions. And your. The health insurance company will also have a lot of different clients. So in this room context, you can imagine that there is a room that spins up for every single RFP that the health insurer needs to respond to. And the agent and human teams are able to access the appropriate room that's scoped to, you know, the data that they need to have access to to deal with the request from that particular client or potential client.

Tula Masterman [00:07:49]: And then as you're going through all of this, you'll also notice in our Mesh agent studio, that's a, that's one of the environments that we provide where you can test this, but we automatically have logging and observ baked in. So that way you know exactly what tools the agents have called, what messages they've sent, how long things are taking. You know, all of that is is tracked in law. So that way you have a really good understanding of not only the final product and output that the agent is producing, but what they performed every single step of the way. So by the time you get to the end, you not only have an output that you know you can, you can trust and that your team had input into, but you also have that audit trail of activities as well. And I'll just kind of show you this. That way you can get a sense of what things look like inside of the studio. With Mesh Agent, we have our SDKs in a variety of different languages.

Tula Masterman [00:08:46]: So we have our Python SDK. We also have a TypeScript, JavaScript and Dart SDK because we want to make sure that, you know, developers are able to build agents and agent applications and the languages that they know best. But as you're building agents, one of the things that we found is you might not know exactly what like your end state application looks like, but you still need a really, really good place to test this information and to evaluate your agents and understand how well they're performing. So when you're in the build phase, we have the Mesh Agent studio, which is what I have a screenshot of here. And this just kind of shows you at a high level. Like, you can see I was a participant who was in the room. I also have a voice agent and a chat agent that are there. Each of these agents are equipped with a variety of different tools to help perform actions like writing documents or saving information to storage that exists in the room.

Tula Masterman [00:09:44]: And I can simply have a chat with them here. If there was another teammate that's in the room, you can just message them directly as well. So you can see that you can communicate with all sorts of participants and then those shared documents just automatically appear in the room. So everybody has an understanding of what's going on and what activities have been taken or what deliverables are being drafted within that shared context. And if you want to learn more, I definitely would love if you reach out to us, you can find me on LinkedIn just at Tula Masterman. You can also find me through my Gmail, which is just Tula mastermanshagent.com and then you can also sign up today through meshagent.com and we have a lot of different docs that are on our Doc site, which [email protected] as well. So all of those, I would say, are good resources to kind of get started and to learn more. And if you're interested in collaborating or you want to build with MeshAgent, please reach out.

Tula Masterman [00:10:44]: We are super excited about the future of collaborative AI and hope to build it with the MLOps community.

Adam Becker [00:10:52]: I. And we are stoked to collaborate with you on this future because this is honestly, like, right before I joined this meeting, this conference, I was trying to debug something that looks exactly like this problem and you're trying to solve. So let me ask you just a couple questions. Do you have any. Are there restrictions on the number of rooms?

Tula Masterman [00:11:16]: No rooms. Those scale up and down automatically, so you can kind of have that configured. So, like, let's say you have one day where there's a lot of activity going on and you need kind of the same set of agents, you know, within each room. You could have like. Like 10 rooms running. You could have thousands of rooms running, just as long as there's participants in there. Or if you just want to leave the room open, you could configure it that way, too.

Adam Becker [00:11:38]: Yeah, yeah. And then is there a way to then chat with those agents in a way that doesn't influence their memory?

Tula Masterman [00:11:48]: That's a good question.

Adam Becker [00:11:49]: You know what I mean? Kind of like.

Tula Masterman [00:11:50]: Yeah.

Adam Becker [00:11:51]: Have you seen. What was it called? Like, West. What was that? I don't remember. Westworld.

Tula Masterman [00:11:58]: Westworld, yeah.

Adam Becker [00:12:00]: You can go in and almost reset. Just actually ask them a question and then you put them in a different mode. Because if I want to go and investigate what happens with a particular agent, execution I might not want to influence.

Tula Masterman [00:12:16]: Yeah, you can, definitely. Our goal is that agents are very configurable. Let's say you want to implement memory in a certain way. You can build the agent to. To say, you know what, like, we want to deal with memory in some way to where we just kind of flush the system or, you know, every time that I come back to this room, like, I want the memory to be persisted. You could, you know, save it and load it that way. Obviously, you're still dealing with, like, context, window constraints and things of that nature, but ultimately, yeah, you can kind of handle memory the way that you want to.

Adam Becker [00:12:47]: Let's see if we have any questions from the audience. Jyoti, do you integrate with other CN ABB tools for a single pane of glass for OPS teams?

Tula Masterman [00:12:57]: For OPS teams, yeah. So we also. We do integrate. Like right now all of our like telemetry is running through open telemetry. But the goal with Mesh Agent is like, if there's integrations that you need to bring in, you can bring those in. If there's other observability features that you want to track, you could have, you know, custom like logging and anything of that nature as well.

Adam Becker [00:13:18]: So yeah, nice.

Tula Masterman [00:13:20]: Customizable. Yeah.

Adam Becker [00:13:21]: Do let's stick around the chat in case folks have more questions. I have a tab open. I just reached out to you on LinkedIn. This is fascinating. I'd love to connect and to keep this going. I believe the future is going to be very multi agent and I also think that questions around memory are going to be increasingly relevant here. And next we have Ben, who's going to tell us us a little bit more about that. So Tula, thanks for joining.