AI Snippets #1
Your weekly Microsoft AI newsletter
đââď¸Hey friends,
Christopher here with another edition of Microsoft AI Snippets! As always, I wont cover every single update, as there are simply too many. Instead, I focus on the key highlights for you.
Todayâs content:
Microsoft AI Updates
Feature of the Week
Learning & Community
Microsoft AI Updates
đ¤ M365 Copilot
GPT-5.2 Model Selector in Copilot Chat
Some of you may recognize this option from ChatGPT or Claude and it works exactly in the same fashion.
You can now pick GPT-5.2 directly in Copilot Chat. There are two modes. âQuick Responseâ gives you fast answers. âThink Deeperâ takes more time but handles complex reasoning better.
The problem this solves is simple. Before, you got one response mode. Sometimes that was overkill for quick questions. Other times it wasnât enough for tricky problems. Now you have more control which is a good thing.
Source-Specific Filters in Copilot Search (GA)
Copilot Search now lets you filter results based on the data source youâre searching. These filters appear automatically when you drill into results from a specific connector. For example, if you search Azure DevOps, youâll see filters like âArea Pathâ and âAssigned To.â
The problem? Enterprise search often returns too much. Youâd end up rewriting queries or scrolling through pages of results. Now you can narrow things down without changing your search.
Even though this is a rather subtle change, it brings practical improvement that actually matters day-to-day. The filters adapt to each connector automatically which makes it easier to use efficiently.
Publish Microsoft Foundry Agents to Copilot
You can now publish agents built in Microsoft Foundry directly to the Microsoft 365 Copilot Agent Store. Your agent shows up in Word, Teams, and other M365 apps.
The problem this addresses is the gap between building an agent and actually shipping it to users. Before, youâd build something cool in Foundry and then spend time on integration work. Now that friction is gone.
This is a big deal for scaling AI solutions across an enterprise. You can now actually build once and deploy everywhere.
đ§ŞCopilot Studio
Copy Agents from Agent Builder to Copilot Studio
Agents you create in Agent Builder in the M365 Copilot app can now be opened in Copilot Studio. This unlocks advanced settings such as better governance controls or analytics which are not available in the Agent Builder in M365 Copilot.
Agent Builder is great for quick prototypes. But it is limited intentionally. If you suddenly wanted more control or proper governance, during the design of the Agent you were stuck and had to start over in Copilot Studio.
Now you can start simple in M365 Copilot and migrate your agent to Copilot Studio to get the full toolset if required.
This creates a natural path from experimentation to production-ready agents, which makes the development process easier. I like it.
đď¸ Foundry
GPT-5.2-Codex Now Available in Microsoft Foundry
OpenAIâs GPT-5.2-Codex is now generally available through Azure OpenAI in Microsoft Foundry.
It has a 400K token context window (roughly 100K lines of code), supports 50+ languages, and can reason over code alongside images like architecture diagrams and UI mocks.
Real enterprise codebases are messy. They just are. You always have some legacy systems, partial documentation, complex dependencies, tribe knowledge and what not.
GPT-5.2-Codex can hold more in memory and reason across longer tasks which will help during those more complex tasks. The 400K context window is genuinely useful for enterprise scenarios. Being able to reason over diagrams alongside code is interesting too.
What I appreciate most with the integration into Foundry is the ability to run it natively on Azure infrastructure with existing security and compliance controls, landing zones and all the cool stuff you can build there.
Im still curious mid to longterm on whether GPT Codex or Claude Opus will have the advantage in enterprise environments. But thats a topic for another week.
đď¸ Feature of the Week
Conditional Access for Agent Identities
Microsoft Entra now supports new identity types specifically for AI agents:
Agent Identity â like an app/service principal but specialized for AI
Agent User â an identity that behaves more like a human user
Agent Blueprint â a template used to create agent identities
You can apply Conditional Access policies to these agent identities, although it is currently more limited than what you have available for human identities in conditional access.
What works today for agents? You can block specific agents from acquiring tokens. You can target agent identities in your policies. Thereâs also an early preview of agent risk scoring. Thatâs basically it.
What doesnât work? MFA, device compliance, location conditions, session controls, authentication strength. None of the controls that require human interaction or device context apply here. Makes sense when you think about it. Agents donât tap a phone for MFA.
This is exactly what Zero Trust for AI should look like. Start with the basics. The current implementation is limited, but it is a big step in the right direction.
Learning & Community
đ Learning Resources
đ Free Microsoft Assessments
đ Free Upcoming Microsoft Events
And thatâs it for this week. I hope you enjoyed it and look forward to the next one already.
Cheers,
Christopher