Welcome everyone. Welcome to the 3rd session of this series, which is AI ops reimagined. We will talk about LLM, we'll talk about and we'll also talk about next.00:09Dashboard MCP server and also MCP server. Today we have Anshul with us who will be our specialist who will take care of all the live demos with us and I will be taking care of the presentations. We will be also having few more panelists who will be joining us shortly.00:20Will be there to answer your questions throughout this session. Now, when you were joining this session, probably you would have seen the Slido that would have come come up.00:41On your screen. If in case someone just joined and were not able to join the Slido, you can take this opportunity to join the slido. You can either do it by entering or logging into slido.com and then entering this code of hash.00:512920292 and then just add the passcode of AI ops or else you can just scan this QR code with your mobile devices and then just enter the passcode of AI ops. And then you will be right into slido. This Slido will be there throughout the session to understand a little more.01:07About your environment, about yourself and how comfortable are you with this new terms of AI.01:28LLM MCP, and how well worst are you in your environment especially while using it, so that we can understand what you're using and we can take up the session accordingly. We will also be coming to this QR once we start the slider as well. So in case you're not able to.01:36Join now, you can again join when we have the questions. But before that, let us move on to the agenda.01:56Now the agenda today looks something like this. 1st, we will be talking about the foundations. This will build the vocabulary to be honest. You will we will be talking about LMs we will be talking about rag, what is tool calling.02:02MCP and everything will be introduced here. Second will be Cisco MCD servers where we will show what exists today and something that is also coming up in the future that we will be highlighting. We'll be also talking about hosted versus community MCD servers and we will also have two demos that will be part of.02:17This. Then we will move on to the 3rd part of today's segment, which will be autonomy versus agent.02:37Which is a conversation talking about honest autonomy and ending with agents, how agents talk to each other and they interact with each other. And finally, we will move on to where is it heading. We'll talk about some of the new.02:44Products that are coming in like cloud control, AI canvas, Cisco IQ, and what you can do on top of that. And in case we are not covering something that you are expecting, you can put in your questions in the chat and we'll try to answer them as much as we can. But now.03:00We have the 1st slider question. Now this is an.03:19Interesting question, that is what is the biggest hurdle in your team for adopting AI agents? And you can put in your votes and we will be able to see how it does it look like. You can scan the QR, you can put.03:22Your questions, your responses through that. Okay, I see we have a mixed kind of response. Okay, five mixed to be honest, we have 25 % in each related to ok now it's updating, you see around 40 % having issues with integration.03:40A lot of people having issues with security and data privacy definitely a trusting AI accuracy, again a valid point where we are still not able to fully trust AI and we still have the concerns whether it will be able to give the right outcome every single time.04:00So that is definitely a valid response. Okay, I see a few more people.04:20Putting in security and data privacy. Yes, as we are moving along in this journey, we are introducing new AI, the security aspect of it also becomes a bit challenge. So we will see if we can talk a little bit about how you can mitigate all these issues, and especially the top issues that we have.04:27Okay, one good thing to see is NO one has selected lack of policy or governance, so I think everyone is moving towards the governance of AI regardless of which industry you're part of, but that is one thing that we can see out of it. But again, most of the people are choosing security and data privacy, and this is just the 1st question and we are.04:47Giving people time to warm up but in the next question onwards we'd like to see more people putting their vote so that we can understand your scenario.05:07Now, let's get started with the session. This slide covers the four different errors that you can see. Let me bring up my pointer as well. So we started with quick ops, then our CLI ops, we started with CLI ops, then we moved to click ops, then DevOps, and now finally we are at AI ops.05:16So in each stage you can see what are the things that are included, so engineer of course we all will be there regardless of which area we are talking about, but then engineer interacts with CLI and then the device. Similarly in clickops, engineer reacts with GUI and then the device. And it was more of inclusion.05:35Of coding or templates and that would be going on to a controller and with the controller in turn will be putting the configurations or controlling the device that you are trying to manage at the end.05:56But when we talk about AI ops, the engineer actually puts in words in natural language. It's not something that you are putting in any instructions in a particular format. You can just talk to.06:08It in normal words and that will go to an agent which will also be integrated with an MCP. Now this agent and MCP combo will will talk to the controller which you can see in the device in the dev OPs as well, and the controller will in turn talk to the device. So this is the new workflow that we.06:24Have the input changes a lot. We do not need to have very strict regulations on what your input would look like. You can just talk inwards and then.06:45The work will be done on the device. So currently we are in the mix of all these, but we are slowly moving in towards the aiops.06:54Okay, now that we have the current scenario locked in, let us move on to the foundations. So what NLLM actually gives you? So this is a.07:11Starting point which we would like to highlight. So what ILLM is good at, it is good at summarizing incidents. It is good at generating hypothesis, translating intent into structure, all these things LLM by itself is good at. But what it is not good at is in knowing a current state.07:25Bit of an infrastructure or a device. The current state is not something that LLM by itself has access to.07:44Or it cannot guarantee accurate unaided responses or deterministic readability. All these are something that it is not good at by itself. Therefore, the model has NO idea on what your fabric looks like right now. It only knows its training data. This this single.07:52The gap is why the rest of the session exists for your knowledge or MCB for the live state, the model choice.08:12Is also per workflow and not one model can do everything. So you need to also choose your model a little cautiously. Now, this is the 1st talking point, a thing called a flag or retrieval augmented generation. This integrates with your LLM.08:21So that it can get a little more relevant material.08:40To you and then ask the model the answer to it. So what happens in this workflow? You ask this question, e.g., what is our spine leaf MTU standard? Then it retrieves information from design documentations, from notebooks, from past cases in case.08:45There are some incident cases on these and then it retrieves that information and also adds to your question.09:05The information that it retrieves, it adds to your question and then builds the whole context. So your question plus what it retrieves from the documentation that you have already uploaded, it adds to the context and then it works on that whole prompt and then generates the response to you.09:12This is the whole workflow that how it looks like, and after the generation it gives the answer to you. This is the five steps that you can see on screen as well on how this entire process looks like. So what it fixed, it fixes that the answers are grounded in your standards and it is not just public internet info.09:30The citations are there, which you can check so that you can ensure that whatever it has referred is actually correct and new documentations can be uploaded as you move along so that the information can be updated as you move to the future. But.09:51Where it stops is it is only retrieval, so it is as good as you have indexed it. It reads document, but it cannot see the fabric. Of course, it only sees static documents and stale pages produces confidence.10:08And it might also give you confident answers when it is outdated. So you only you need to keep updating it with the new information with the new documentations as you move along. Now the bridge that comes into this is.10:24The from chat to tool calling the bridge that we're talking about, there are four steps to it. So the 1st step, the user asks the question and then the LLM will select the model on which you want to work with. The model will execute a tool and then the LLM will summarize.10:42With evidence. So the model does not run the command, but it chooses which command to run. That's all.11:01So when you compare the two panels at the bottom, you see that the order itself gets this kind of an out in input and then the tool in turn runs this command on the device. So this is how a tool comes into the picture where a user gives a command.11:08In national language, the tool understands which or the LLM understands which tool to run and then the tool gets done by the LLM and then finally the LLM gives you the response with result and evidence as well.11:28So now we come to MCP, as we were talking about tools, the LLM can have access to one single tool, but if you want to use a multiple number of tools in this case, then you would be better with an MCP. So what it looks like without MCP is.11:43In case you have an agent, it has to access all these tools separately. So it has to decide on which tool to use on its own. So you see there are into any kind of interactions, so the number of interactions increased by a lot. The LLM has to go to each number of tools and then.12:02The increase, there is a lot of increase in interaction from the LLM side. But with MCP, the LLM only has to interact with the MCP. MCP will in turn help the LLM to understand which tool to use and then with that together.12:20You can see that the number of interactions with the LLM produces a lot. So in the left diagram, this is without MCP. The right diagram is with MCP. So this is like a USB C analogy. It is about eliminating the bespoke glue.12:36Nothing more, nothing less. If the room has one definition, so this is the recommended one that you should go ahead with so that the interactions with the LLM itself reduces and that also reduces your overhead as well.12:52Now, what happens in an inside an MCP connection? So there are two halves to it. Each is owned by a different party. So the 1st left hand side is the client side. So you can talk about cloud desktop cursor or any assistant that you have on an audio to own agent. So this holds the LLM integration.13:09So the client is actually the part that holds the integration with LLM.13:29And it also orchestrates the conversation. It maintains the connection as well. Now if you come to the server side, this is where the services or the tools that are there that is integrated with the MCD server, the tools that will interact with the devices or make configurations or get data from the devices, these.13:34Tools are on the server side and they are maintained by this side of it.13:54So the server is where your governance lives, the scope, the authentication, the audit sit there and the model interacts with the client side. The transportation is either over HTTP or it could also be locally on your environment. The choice is usually a security conversation.13:59Session that which one you want to use. If asked.14:18In case the maturity here, the specification quickly moves from one to another and there is interaction between the client and the server and this all works together inside the MCP. So this is how MCP works. We have the client side, we also have the server side.14:23Okay, now let us move on to Cisco MC servers, what are present from Cisco itself, which we will try to cover in today's session. So we have posted.14:44Servers and also community servers. So there are five different dimensions to judge any MCB server which are highlighted here, the ownership, the support, security cadence, and this part, so each of the two which is hosted or community are mentioned in this way. So in case of ownership, you can.14:57Can see that a hosted MCP is vendor managed so it is fully hands off. You do not need to manage it yourself, but a community will be managed independently and in a maintained repository. Support, backend support and.15:17Back and support path and defined life cycle is there for a hosted MCP, but if we talk about a community MCP it is based on best effort as much as possible, we try to give support because it is maintained by a community. Security.15:35There is central review and enterprise governance in case of hosted MCP, but if we talk about community, it varies, it depends on how the assessment is done or how it is developed, so it depends on that cadence aligned to the product release cycles, there are defined cycles on which.15:50There is a cadence of updation of the hosted MCP, but if you talk about community, the community drive or the contributor driven, it depends on the contributor how much she or she can contribute to the cadence and updates of it. And the hosted ones are best fit for production environments where there are workflows which.16:10Quick control.16:30And if you talk about community, these are more for pilot cases or if you want to build your own custom workflow or if you just want a fast integration with your environment, you can use the community as well. So this is more for testing or once you are trying to implement it in a test environment and then try out the hosted.16:31You can go ahead with the community one, but again, the hosted one will definitely be a more production ready MCP server. So neither of them is wrong, it completely depends on you on how much you want to experiment.16:51With either of them and then you can choose. But just a rule of thumb would definitely be that posted will be better for a production environment with strict control and community will be better for pilots or if you want fast integration. Okay, now if we talk about what exists.17:09Today we have a hosted our vendor run Nexus dashboard MCP server that is coming in shortly. It is.17:29Not completely available as of now, but it is in the works and will be available soon to you and you can start utilizing it and that will be in case of a hosted or a vendor and MCV server. But when we talk about community one, we have an ACI nexus or ACI MCV server.17:38This is a community server in front of Apic API, which is published on Cisco code Exchange and you can go to the devnet website and explore it as well.17:58So anything with an API is just a wrapper away, so any environment that has an API, you can convert that into a tool and then integrate it with MCP server just with a wrapper. That's all that is required.18:09Now, if, when we talk about both of these, we will be seeing them in action today. Now, when we talk a little bit more in detail, the next dashboard MCP server is a hosted server for day two. So we have, as we talked about MMCP client, we have.18:25Then there are services of next to dashboard like inventory assurance or telemetry and that MCP server integrates with these services and then you get an outcome like insights evidence on next actions.18:47Based on whatever your prompt is. So these are the two categories that we are talking about. We have a fabric inventory, health anomalies, telemetry, configuration contexts, we also have guardrails where we can access the scope by role or we can give access by role, we have approval gates on.19:02Changes that you make. We also have audit trails for every request in case you want to look back.19:22On what has happened on the system and of course you have observability so prompt to tool call is there, you have structured payloads, you also have transparent errors. Now this is the next dashboard MCP server. We will be showing the demo towards the end of this session, but before that we will also move.19:27On to one small question on Slido as well.19:47So this question would highlight on if in case we're talking about ACI because that is the next MCV server that we will be showing and we will also talk about that. So which ACI task would you trust an AI agent to handle?19:51Which of course your final approval. So this also talks about human in the middle or there is these approval gates that you talk about in which the AI will do the work, but it will of course have an approval gate where a human will give the final approval if in case that system is interpreted in your.20:10Environment, which AI task would you trust the AI agent with?20:30Okay, we see a lot of people opting out for troubleshooting and root cause analysis policy configuration, change management, reporting and documentation. We see a lot of people choosing troubleshooting and the root cause analysis.20:35Okay, so I see this pattern where troubleshooting and root cause analysis coming on top, their policy and configurations coming second and reporting and change management coming in as 3rd. So that is good to see that at least.20:56We are ok to let AI come into the picture if in case we have a final approval, of course, which is very important at least still now until we have a very robust autonomous system which does not require human approval, but before that, we will always have a human giving us approval.21:13On what is something that the AI might have reduced to be true, but the human should finally take the judgment whether he or she feels if the approval is given or not. So ok, good to see a lot of people choosing in troubleshooting and root cause analysis. We will try to keep that in mind.21:32And for the future part of this session. Now I can show the NDMCP server for now. Probably we can do the end so I can quickly show the team how it works.21:51Sure, so we had the demo, but there was some glitches in the back end, so we wanted to keep it at the end, but I think Anshul has figured out it so we will hand over the stage to Anshul's to show us the next dashboard MCP to all of us.22:09Sure, thank you thank you but again as always, you know, Murphy's law, things, if something has to go wrong it actually goes wrong just just one of the case so but for fortunately we were able to salvage.22:27Just a bit on MCP server like Ronak was saying MC per server is just nothing more than a gateway on top of your APIs, which exposes a set of tools to your LLM, right? So now I'll show you in action I'm using VS code tool with github copilot.22:44And connecting to the Nexus dashboard to be able to do that. So I'll just show you how it all hangs together. So this is my VS code to environment, and if you look at it, I've got a bunch of MCP servers running, so if I go to my.23:00Profiles, MCP servers, I've got quite a few MCP servers running over here, right? And as you see over here, I've got MCP server, sorry, I can't zoom Beyond, if you could zoom in, people are, I think not able to see it properly.23:16I'm not sure because this is a VS code too if I can zoom in any more than this unfortunately. In case you want to zoom in from your end, you can just click on this plus icon on the Webex and you can zoom it on your end as well. I mean to the audience.23:33This one. Okay yes yes. Thank you. So, as you see over here, I've got a MCP server running, it's called M my NDFC MCP app server, and this is running locally on my laptop, you know, you see this it's just a python script running locally on my laptop and.23:51
