Build an AI DevOps teammate transcript

Hello I'm Vishal Bulbule. I am a Google Developer Expert in Cloud and AI. I'm also a founder of TechTrapture which is an AI startup. So today we are working on a challenge. I just received an email from Google Cloud where the production systems are failing. This is the issue we have to fix. We have to analyze the log. We have to troubleshoot the issues. We have to perform a Root Cause Analysis and we have to create a postmortem report.

Along with that, we have to fix the issue. So instead of doing all these things manually, we are going to build an AI DevOps teammate using Antigravity, Google Cloud MCP Servers and Google Gemini. Let's go and see how we can build our AI DevOps teammate. Let's discuss the issue first. So we have the web server running on a Google Compute Engine and on the web server we have the Apache application running. So currently, users are not able to access the application.

And you can see here, users are getting an error that shows they are not able to access the application. And this is the issue we have to troubleshoot. We have to find the logs and we have to fix it. So we'll be using Google Antigravity, Google Cloud MCP servers and Google Gemini. So we'll be using two MCP servers here. First, we'll use a Google Cloud Logging MCP server. And secondly, we will be using a Developer Knowledge MCP server.

So I am in Antigravity. Now you can see here, we click on the three dots here and we can click on MCP Servers. You will see the MCP Store and you will find the list of MCP servers available. So these are all remote MCP servers. So as we discussed, we need two MCP servers. First we will use a Developer Knowledge MCP server. So here you can see a Google Developer Knowledge MCP server. So this is built on Google official documentation purpose.

So I'll just click on it and install. So it will be installed in our Antigravity. And I'll refresh it and I'll see what tools we have. So here there are three tools. First, search_documents. Whenever we need to search through the Google official documentation, it will send a query and it will use a search_documents tool to go through the official documentation. Second, we have answer_query and finally, get_documents.

So we'll be using these three tools along with the Google Developer Knowledge MCP server. So let me go and connect to another MCP server which is Google Cloud Logging. So I will install this as well. So here we have list_log_entries, list_log_names, get_bucket, list_buckets, get_view and list_views. So there are six tools available. For now, I will enable all the tools. So whatever the agent requires, it will fetch that tool and use it to debug and troubleshoot the issue.

So let's refresh our MCP server list now and check if we can see the tools available. Okay. So now you can see that Google Cloud Logging is available. And a Google Developer Knowledge MCP server is also available. So now I will go to the prompt. Okay. So we are adding the persona here. "Act as a DevOps Engineer and help me to troubleshoot the issue." Next, we add the issue description. So I am adding here: "Users are not able to access the web server running on a compute engine" and I add the VM name.

So the VM name is web_server and I also add the project name tt-labs-02. So I'm just giving more details. So it will not ask any follow-up questions. And it's best practice to give detailed information in your prompt. So it will give you a better response. So the better the prompt, the better the response and the reasoning you can expect from your agent. I am also adding some rules here. "Please analyze the logs, troubleshoot the issue using a Google Cloud Logging MCP server." If needed, you can refer to Google Cloud official documentation.

I am also adding one strict prompt: "Only recommend the fix. Do not make any changes without any confirmation." So it will only provide us a recommended fix or a solution, and it will not go and apply the fix in a production. And that keeps the human element in the loop of our process. Also at the end we are asking it to find the root cause and prepare RCA and postmortem reports. So this is the whole prompt and set of instructions that we are giving to the agent.

So I will just press 'enter' and see our agenda in action. Now you can see it found the Google Cloud Logging MCP server. So it started analyzing the log. And it also found the Google Developer Knowledge MCP server. So it should also have access to the Google official documentation. Now the standard process is to first check the logs. So it finishes the log and saves it to output.txt in the temp directory and it is now analyzing the log.

So it is fetching a few more logs. Now you can see it is checking the logs related to the firewalls and the instance configuration. So it will try to list all possible causes which can cause the issue. So it is checking the instance configuration, it is checking the firewall and it is also checking the logs for any activity that has been made in the project. Now you can see, it is searching for official Google documentation.

It found something in the logs and then it is just validating against the Google official documentation. You can see here it is checking a "gloud compute instance add-tags" command. So it is just trying to see which gcloud command is required to set up the firewall rules. Now you can see it created an RCA report within just a couple of minutes. We created the whole RCA report. We can just go through it first, so you can see the Incident Overview.

Based on this analysis and the description we have provided, the web server on Compute Engine instance is inaccessible. The project ID affected resource is shown, and it shows the Technical Root Cause. So the Compute Engine instance lost web traffic due to the missing network tags. So this is the RCA it found. And it did not only find it in just one line, it dug down into the exact configuration. It dug down and found an ingress rule prod-vpc-allow-http on a port 80 and it lost its http-server tag, and because of that, the VM instance lost connectivity on a web server.

So you can see here this email ID has misconfigured the tag which caused the whole outage on the web server. There was an impact and the logs also captured it. The traffic was denied in the logs. So it also provided all the log evidence which caused the Tag Modification, which also caused the Firewall Ingress Denied Traffic. And it also provided a Recommended Remediation. So as we said in the prompt, do not fix the issue in production and only provide a solution in the commands.

Now we can see here it provided the commands and we will validate it and execute the commands, and we will see if it can fix our issue. Just accept the changes for the file and command to apply the fix. So I will just go here and it is asking for confirmation. I can either give the confirmation here or I can go and execute the command by myself. So I prefer to use the safer option. I'll just go and execute the command to fix the issue.

So what it is doing, is just adding the tags to my web server. So you can see here it is adding two more tags http-server and https-server. And after these tags are added, the issue should be fixed. But we will add the tags and then validate if the agent has provided the accurate fix. Now the tags are updated. So let's go back to the VM and just reload and see if we can access the web server now. And now you can see, we can access our web server.

Whatever fix that Gemini provided is working accurately and our production issue is solved with a single prompt. And finally we fixed the issue. Let's summarise what we did. So I'll just go to IDE and go through what we did one by one. So initially based on the issue description, our prompt contained the detailed issue description. We provided the VM name and the project name. And based on our description, it went to a Google Cloud Logging MCP server.

First, it analyzed the log. It saved the log into output.txt. After that, it just verified the multiple logs like Google VM configuration log, the cloud logging activity log and it also checked the firewall logs. After that it performed a Root Cause Analysis and then it went to a Google official documentation using a Developer Knowledge MCP server. It found some documentation and it created the recommended fix. So it gave accurate commands using the Developer Knowledge MCP server, and at the end it created the RCA report and postmortem report.

We also checked if the recommended fix is working, and we applied the recommended fix, and our web application started working properly. And that's it. The challenge is completed. So based on the issue description we just created one simple prompt. And we then provided the prompt to the Antigravity agent. Within Antigravity, Gemini used a Google Cloud Logging MCP server to analyze the log. After that it went to the Google Knowledge MCP server, where it checked what were the commands required to fix the issue.

I started using AI not just like a chatbot, but I started interacting with it as my DevOps teammate. It wasn't only the summarisation of the log that impressed me. What impressed me, was that it took all the steps that a senior DevOps engineer would. Now you can integrate the MCP servers in your agentic workflow and you can integrate it completely with your infrastructure. It can take action on behalf of you, but you can still have control, keeping the human element in the loop within your workflow.

And you can use all these MCP servers provided by Google Cloud. I have added all the links in the description. You can check the links and try yourself.