Hi there Of us!
If you happen to sat by this session in the course of the Microsoft Azure Infra Summit 2026, you already know that Anand Guruswami and Arun Rabindar from the Cloud Native Experiences group confirmed us one thing I’ve been ready to see for some time. An AI agent that doesn’t simply spit out a Terraform file from a obscure immediate, however truly thinks about your workload, talks to you about it, after which arms you one thing you possibly can put in entrance of a pull request reviewer with out holding your nostril.
That is the Azure Deployment Agent, and on the time of broadcast it was nonetheless in preview inside Azure Copilot, with the identical brains transport as an open supply talent you possibly can plug into GitHub Copilot, Claude Code, Cursor, or no matter your group makes use of. On this put up I need to break down what they confirmed, why it issues for IT execs, and how one can get arms on with it.
📺 Watch the session:
Allow us to be trustworthy in regards to the daily. More often than not we aren’t constructing a model new workload from a clean canvas. We’re stitching assets collectively separately, copying patterns from a earlier venture, searching down the precise SKU, checking quotas, then arguing with coverage on the way in which out the door. Completely different admins do it alternative ways, and that inconsistency is the place threat lives.
Here’s what the Deployment Agent adjustments for us:
- It strikes the dialog up a stage, from “which useful resource do I click on” to “what am I truly making an attempt to construct.”
- It grounds the structure within the Azure Nicely-Architected Framework, so the output is just not a generic LLM guess, it has reasoning behind it.
- It separates the plan from the code, so that you and your group get to evaluation structure earlier than any Terraform or Bicep will get written.
- It plugs into the instruments we already use. Azure portal for the guided path, GitHub Copilot and Claude Code for the ability consumer path.
Briefly, it is about taking the boring repetitive elements off our plate so we are able to deal with the elements that want human judgment.
The Deployment Agent is a functionality contained in the Brokers (preview) expertise in Azure Copilot. Consider it as a digital cloud answer architect that lives in your Copilot chat. You describe the workload in pure language, and it walks you thru a multi step course of to land on a manufacturing prepared deployment.
A couple of issues that stood out from Anand’s portion of the session:
- It helps multi flip dialog. You may make clear scale, safety posture, resilience, SKU preferences, area constraints, and the agent will fold these into the plan.
- It produces a human readable infrastructure plan first, full with commerce offs and the reasoning for every useful resource selection, earlier than it ever writes infrastructure as code.
- Immediately it generates Terraform contained in the portal, with Bicep assist touchdown within the portal expertise shortly. Within the GitHub Copilot movement you possibly can already decide Bicep or Terraform.
- As soon as the plan is authorized, you get an actual artifact. You may open it in VS Code for the Internet, or have Copilot open a pull request straight into your GitHub repo.
The deployment itself nonetheless goes by Azure Useful resource Supervisor. That’s essential. Your tenant insurance policies, RBAC, naming conventions, and present guardrails all nonetheless apply. The agent is just not bypassing your governance, it’s producing code that flows by it.
Arun did an excellent job pulling again the curtain on the internals. The agent follows a two step sample that provides you management at each checkpoint.
- Intent seize. The agent takes your immediate and clarifies the scope, the constraints, and what success seems like. No guessing, no leaping straight to YAML.
- Plan technology. It produces a structured infrastructure plan with inputs, sub targets, a full useful resource record, configurations, SKUs, and a per useful resource reasoning part.
- Validation in a loop. The plan runs by evaluators backed by the Nicely-Architected Framework pillars (reliability, safety, price, operational excellence, efficiency effectivity). If one thing fails, the agent regenerates and tries once more till the plan is strong.
- Human evaluation. The plan is introduced to you in plain language. You may iterate. You may say “prioritize West US 2,” or “swap that SKU,” and the agent will replace the plan in place.
- Code technology. Solely after you approve the plan does the agent emit Terraform or Bicep. The generated code goes by syntactic validation as properly, once more in a loop, so it truly parses and is able to apply.
Below the hood within the GitHub Copilot and Claude Code path, the group has decomposed all of this into an open supply talent (the Azure Enterprise Infrastructure Planner) plus the Azure Nicely-Architected Framework as an MCP device. The bottom agent in your editor picks up the talent, runs the phases, calls the MCP device to floor the output, after which writes the IaC. Identical workflow, totally different host.
This isn’t only a toy for greenfield demos. A couple of locations the place I see this paying actual dividends:
- New workload bootstrapping. A group wants an online app, SQL backend, secrets and techniques in Key Vault, monitoring, and a sane area technique. As an alternative of three days of clicking and duplicate pasting, you describe it and evaluation the plan.
- CSV ingestion to SQL automation. The Claude Code demo Arun ran was precisely this. CSV lands, will get processed, rows replace in SQL. The agent picked smart assets, justified every one, and produced Bicep able to commit.
- Standardizing throughout groups. Completely different admins ending up with totally different shapes for a similar workload is the silent killer of operational consistency. A shared agent with a shared planner talent drags everybody towards the identical Nicely-Architected baseline.
- Ability leverage for smaller groups. Not each group has a deep Azure architect on employees. The agent encodes numerous that have and surfaces it as dialog.
- Open supply customization. As a result of the talent and MCP tooling are open, platform groups in regulated environments can fork it, add their coverage context, their tagging guidelines, their naming conventions, and ship a tuned model internally.
One trustworthy tradeoff. Proper now the agent is greenfield first. The group is actively engaged on brownfield eventualities, pulling insights from present workloads and referencing present assets. If you happen to stay fully in a fancy present property, anticipate the expertise to maintain getting higher over the following couple of releases.
If you wish to strive it this week, right here is the quick record:
- Ask your Azure tenant administrator to allow Brokers (preview) in Azure Copilot. The toggle lives within the Azure Copilot admin middle, and with out it you’ll not see agent mode in chat.
- Within the Azure portal, open Copilot, increase to full display, and swap on Agent mode on the backside of the chat panel.
- Describe a workload in plain language. Be particular about area, scale expectations, and any compliance constraints you care about.
- Overview the generated plan earlier than approving. Have a look at the commerce offs part, that’s the place the agent reveals its work.
- For the editor path, set up the open supply Azure Expertise plugin from the microsoft/azure-skills repo, level your IDE on the Azure MCP Server, and run the identical workflow inside GitHub Copilot or Claude Code.
- Ship suggestions. The group is transport quick and the roadmap (brownfield assist, reference workloads, scoped agent permissions, richer structure diagrams) is formed by what you inform them.
If you happen to loved this session, the total Microsoft Azure Infra Summit 2026 playlist is up on YouTube. Classes on Deployment Stacks, the SRE Agent, Azure Native, AKS networking, and much more are all in there. Bookmark this one and share it together with your group: https://aka.ms/MAIS/2026-Playlist
Drop your questions, your warfare tales, and your want record for the Deployment Agent within the feedback. I learn them, the product group reads them, and your eventualities are precisely what shapes the following preview drop. What would you construct with it first?
Cheers!
Pierre Roman
