Bart Farrell: First things first, for people who don't know you, who are you, what's your role, and where do you work?
Rashad Sirajudeen: Hi, so I'm Rashad Sirajudeen, and I'm a software engineer at WSO2. I'm also a maintainer on OpenChoreo. So at work, I mostly work on AI-related stuff. So Rayson loved me working on the platform agents in OpenChoreo, and also the agentic developer experience, such as the agent skills and the MCP servers and so on. So that's about me, and it's good to be here.
Bart Farrell: All right, Rashad. Dope. How can AI agents leverage OpenChoreo's observability capabilities to diagnose the root cause of Kubernetes incidents?
Rashad Sirajudeen: so like when it comes to AI agents in OpenChoreo, like we have two different classes. The first is the inbuilt agents that we have in OpenChoreo itself and others are the external agents such as Claude code, Codex, and so on. So like when designing the OpenChoreo and the tool servers of our MCP servers, the main focus of us was that the agents are also first class users of our platform which means that the agents also reason through like the same mental model and the abstractions that a platform engineer or the developer uses to like deploy the applications on OpenChoreo so what that means is that the agents so it's true that OpenChoreo runs kubernetes underneath but it does not mean that the agents has reason through that so they can rather like follow the whole OpenChoreo model and then so it has the mental picture of the whole topology and it would reason through that and also we do have two MCP servers so we do have the OpenChoreo Control Plane MCP server which is basically like it gives an overview of all the components, the services endpoints and their dependencies resources and so on and so it would use that to like get a good idea on like what are the services out there and like which they depend on and then we do have our observability MCP servers and we could actually have many like depending on how many observability things that you have so using this it can pull out telemetry of like different components and so on so what this means is that like the agent has like a very good idea on like how all the pieces fit in together and it can like also look at all the telemetry such as logs metrics traces and so on so like it was the same way like for external agents and intelligence has failed but like benefit of using internal agencies that like for example we do have an sr agent like which does ask you on like any alerts that you set it up on so what this means is that like let's say there are some alerts which fire then so when a human comes to the computer and like he looks at it like he has like a whole summary of like what went wrong and where the agent looked and like maybe like what are the things that we can rule out and so on so we do have that so and also fewer like the developers use maybe Claude or Codex, and they can use that to develop their application. And without doing any context switching, they can connect to our MCP servers and deploy it there. And again, if there's something wrong, they can do the remediation themselves.
Bart Farrell: How can OpenChoreo enable AI-assisted remediation? And where should human oversight remain essential?
Rashad Sirajudeen: so again in here, I'll take you through the two different agents. we do have the internal sr internal sr agent which is like so we set up a few alerts and like when an alert triggers like the agent acts on it so it is intentionally read only which means that there are like no any tools exposed to the agent which can do any write operations so that means it's like a hundred percent safe but like there are cases where like the agent researches throughout like different things and like it figures out okay like this is the particular fix that I had to do but it won't go ahead and do it so at that point it would hand over to the user and then the user would review it and then he approves and only then like it would be applied so like here we do have the safety and like we can like make sure like it won't take any destructive actions on its own without like a human oversight And also, there's one more thing that we really like. focused on which is for example let's say like in OpenChoreo what like the main deployment model is that we just build the workload once and we can promote it across environments so let's say there's an alert which triggered in the staging environment and then like when the agent reasons across that and like so there are for example we have a we have an abstraction called the component like and we can promote the same component across different environment so like if the agent is from on this. staging environment and if the fixes from the component level it won't like suggest to go and change that because then that means that it would affect the other environments as well so like we intentionally scope the whole investigation into like a particular environment and a project so which means like it stays within its own limits and it doesn't give us any surprises like on if something changes and also like if it's gonna go away and like do any changes like it would always like wait for human app code so
Bart Farrell: Rashad, how can platform teams enforce security, governance, and operational guardrails as AI agents gain access to production environments?
Rashad Sirajudeen: so let me start on security. So like when you configure the MCP servers to work with agents, so first of all, we have to give it an identity. So in OpenChoreo, we do have an authorization model where we can select for a particular user or group, like what are the things that we allow it to do. like especially for production like when we are deploying platform agents like we have to be careful to okay like give only the specific permissions that the agent is allowed to use and also for my mcp server side like we do have additional optional checks such as for example like when you connect to the mcp server and like get the tool list and like depending on what token and what permissions you have like it will filter out the tools that you don't have access anyway so like we block it off and from that state so like that's one thing that we do and also like OpenChoreo as a platform like gives different abstractions and the golden path like to get things done like on an internal platform and so this is like the same for both the agents and you so that means that like the agents would interact with the platform in the same way that a human would do which means like it should also follow the same best practices and like the golden path so it's like regardless of who uses the platform, it's the same thing that's exposed to them. So we have that. And also all the APIs as well as the MCP tool actions are audited within the platform. So in case something happens, we can trace back to analyze what went wrong and then maybe we can take action on what to go ahead with next. And also these agents are, let's say, Let's talk about our internal agents. They are always mostly scoped to a particular environment or project, which means even if it's a production environment, it would not go beyond that and it will only act within where it is supposed to act on. So we do have that. And also, one main important thing is Open Courier runs on top of Kubernetes, which is you And so if we allow the agents to work in the Kubernetes layer, it's really easy to mess up because we have different CRs and components and dependencies and so on. So even for a human, it's hard to understand. But since we give a view of the platform through the same abstractions, that means that the agent doesn't have to deal with Kubernetes at all. But also if it really needs to, it can dig deep in. So that's where our skills come into play. We are like, OK, so we do have our MCP servers, but still there's lots of domain knowledge which is not available to the agent. But by using our MCP servers and the skills, they complement each other really well. And that allows us to make agents first class users of the platform and get things done. Probably for like anything that we have. So,
Bart Farrell: Well, Rashad, thank you so much for sharing your time, knowledge and experience with us today. Look forward to speaking to you soon. Take care.
Rashad Sirajudeen: You too. Thanks for having me.
