The starting point was a simple definition - agent = model + tools + context - and the loop it runs in (think → tools → observation → response), triggered by a human, a cron, an event or an API call.
The main takeaway: an agent is just a microservice - a regular application or an event-driven worker. Its state is a database, while tools and the LLM are simply APIs it calls. From that angle, “where to host an agent” is the same question we ask about any other service:
An important note about frameworks such as Microsoft Agent Framework: they are building blocks, not a finished product.
Logic Apps got a lot of attention, as they can act as an A2A server, an MCP tool, a knowledge base, or a workflow calling other workflows.
A separate thread was web search. The Bing Search API is effectively frozen and the only way forward is Grounding with Bing Search, which is really just a billing object. Worth remembering: data leaves the Azure compliance boundary, the service works on a cached copy of the internet, and it is hard to debug because you cannot see what goes in and what comes out.
The closing punchline - the Death Star is back, this time with an LLM - about the risk of building one huge monolith around agents again.
The second talk was about automating Azure subscription creation, so that teams can get subscriptions in a self-service way while staying within organizational standards. It was demonstrated with ChrisPolewiak/AzureSubscriptionVending.
The solution is an Azure DevOps pipeline that validates inputs, bootstraps a resource group and a user-assigned managed identity with Bicep and Azure Verified Modules, configures RBAC, optionally creates a service connection with OIDC authentication, and finally moves the subscription into the target management group.
More on the topic:
The third talk - 5 lat czy 27 dni? Disaster Recovery za rozsądne pieniądze… by Kamil Mrzygłód - I unfortunately had to miss.
]]>The exam targets DevOps engineers, developers, and IT professionals with intermediate experience in GitHub Actions – authoring workflows, automating processes, and maintaining CI/CD pipelines.
The exam itself takes 100 minutes, is proctored, and requires a score of at least 700 to pass.
What helped the most, though, was hands-on practice: writing my own workflows, debugging the ones that broke, and building reusable workflows shared across several repositories.
]]>| Service | Tool | Description |
|---|---|---|
| Storage (Blob / Queue / Table) | Azurite | A lightweight emulator written in Node.js. |
| Cosmos DB | Azure Cosmos DB emulator | A local NoSQL database with Data Explorer, also available as a Linux version in a container (vNext). |
| Service Bus | Service Bus emulator | Queues and topics run as a Docker container. |
| Event Hubs | Event Hubs emulator | Event streaming locally, with support for the Kafka and AMQP protocols, run as a Docker container. |
| Functions | Azure Functions Core Tools | The full Azure Functions runtime run locally, with triggers and bindings. |
| API Management | Self-hosted gateway | An APIM gateway running locally / in a container. |
Besides the emulators listed above, it is also worth taking a look at LocalStack for Azure, which lets you test infrastructure code without physically deploying it in the cloud. Unlike the other entries, however, it is not a fully free solution — running the emulator requires an account and a token (LOCALSTACK_AUTH_TOKEN).
So there are quite a few options for learning and testing the infrastructure you are building without paying for the cloud — it is worth using them.
]]>The assessment covers hands-on skills in:
Applied Skills assessments are based on interactive labs – instead of answering theory questions, you perform real tasks in an Azure environment. That makes them much closer to actual work than a traditional exam.
]]>The assessment covers hands-on skills in:
Applied Skills assessments are based on interactive labs – instead of answering theory questions, you perform real tasks in an Azure environment. That makes them much closer to actual work than a traditional exam.
]]>Ingress — more expressive, extensible, and developed by SIG Network. Unlike classic Ingress, it separates roles clearly (GatewayClass, Gateway, HTTPRoute), supports traffic splitting and header manipulation natively, and works the same way regardless of the provider (Istio, Cilium, Envoy Gateway).
I put together practical examples in k8s-gateway-api covering three approaches on a local KinD cluster: classic NGINX Ingress, Gateway API with KinD’s built-in controller, and Gateway API with Istio for full service mesh integration. The whole setup is automated and fits in a handful of commands — a good way to learn the standard without cloud access.
]]>I will keep this list updated as new materials and certification updates become available.
]]>While preparing, I focused on consistency and hands-on practice, not just theory.
I mainly used these resources:
It was a demanding but very rewarding stage. It gave me more confidence working with Azure and even more motivation to keep growing.
]]>VNRA is deployed in a dedicated subnet named VirtualNetworkApplianceSubnet within a virtual network. It acts as the forwarding layer for routed traffic, replacing the need for VM-based NVAs and their associated load balancers. This approach removes the performance ceiling imposed by VM SKUs and simplifies the network architecture.
To understand where VNRA fits, it helps to look at how NVAs are used today. A practical example is described in the article Azure VMware Solution with vWAN Routing Intent and Palo Alto Cloud NGFW, which demonstrates a hub-and-spoke architecture where a Virtual WAN hub with Palo Alto Cloud NGFW acts as a central inspection and routing point.
In that architecture:
10.0.0.0/8, 172.16.0.0/12, 192.168.0.0/16) to all connected networksThis pattern works well but introduces routing bottlenecks when traffic volume grows, especially for east-west flows between spokes. VNRA addresses exactly this problem by providing a hardware-accelerated forwarding layer that can scale horizontally without the performance limits of VM-based NVAs.
VNRA represents a shift toward hardware-accelerated, managed routing in Azure. For organizations running large hub-and-spoke topologies with heavy east-west traffic, it promises to eliminate the scaling challenges of VM-based NVAs while keeping everything Azure-native. Worth watching as it moves toward general availability.
]]>In a recent episode of the On Purpose podcast with Jay Shetty, Novak Djokovic opened up about his mental battles and the powerful mindset shift that transformed his approach to tennis and life. Despite being one of the greatest athletes of all time, Djokovic admits that he constantly battles internal struggles: “I might appear locked in… but trust me there’s a storm inside.”
His key insight? Mental strength isn’t about eliminating negative thoughts—it’s about not lingering in them. As his mental coach taught him from a Zen master’s wisdom: “No one on this planet lives without negative thoughts. The difference is how long you stay in them.”
Djokovic trains his mind daily through meditation, visualization, prayer, and journaling. His message is clear: mental toughness isn’t a gift—it’s developed through consistent work, just like any physical skill.
Why am I sharing this? Because whether you’re debugging code, managing cloud infrastructure, or facing any professional challenge, the mental game is often the hardest part. Djokovic’s approach offers a practical framework for anyone struggling with self-doubt.
I highly encourage you to watch the full episode. The lessons on mental resilience apply far beyond sports.
]]>