Microsoft announced Agent365 in November 2025 and expanded upon it in the Wave 3 announcement a few weeks ago. It's so
This Isn't About Models
When people talk about AI agents, they usually mean "ChatGPT that can also call an API.”
Agent365 is a control plane for enterprise agents. It's the platform underneath the agents, not the agents themselves. Microsoft built it to solve the problem that were starting to develop before it got out of hand: agentic sprawl. Every team building their own agent, no shared identity, no consistent governance, no way to audit what ran, when, and why.
A single control plane fixes that. One place to create, govern, monitor, and shut down agents across the organization.
Three Things That Make This Different
Identity. The January 2026 SDK gave agents a real Entra identity. Not a shared service account, not a shared API key. Each agent gets its own Entra Agent ID with least-privilege, just-in-time access - the same model used for human employees. An agent accessing SharePoint only gets the permissions it needs, for the duration it needs them, with a full audit trail. Defender and Purview govern it the same way they govern people. Agent auth has never worked this way anywhere else.
Native reasoning. GPT-5 integration means agents don't just respond to prompts. They reason about what needs to happen and act on it. You don't wait for someone to ask for the board report. It knows Monday is coming, connects the dots across Dynamics data, and produces the output. The agent is reasoning over a goal, not executing a script.
Proactive execution. Agents run on schedules, triggers, and conditions - not just on user prompts. This is the real power. An agent that monitors your Sentinel alerts and drafts triage notes before your analyst gets to them in the morning isn't a tool. It's a process. One that runs whether anyone asked or not. It also just happens to be essential for cybersecurity these days. Ask our ThreatDefender team about the impact of autonomous agents in the future for security.
What This Looks Like in Practice
The board reports is the clearest example because it's concrete. But the architecture supports any scenario where the work has a pattern: monitor this, detect when something changes, do something about it.
Security operations is an obvious fit. Defender alerts come in. An agent with Sentinel access and the right reasoning model triages them, correlates signals, and drafts recommendations before a human analyst opens their laptop. The analyst still decides. The prep work is done. There are also obvious implications to sales, marketing, and just about any other knowledge worker practice that exists. Finance, HR, compliance, field operations, customer success and any function with repetitive monitoring work is in scope. The agents are doing the work that used to mean someone staring at a dashboard for two hours a week.
What I'm Waiting For
I'm watching two things.
First: how the governance evolves as this area of IT rapidly changes. The Entra Agent ID model is theoretically sound. Least-privilege, JIT, auditable. Whether organizations actually configure it that way or default to broad permissions because it's faster is always a challenge.
Second: how this resonates for mid-market IT teams who are already stretched. The board reports can be a great demo because it's a self-contained workflow with a clear output. Real deployments are messier and take time to structure. Agents that touch multiple systems, partial data, ambiguous triggers, etc. are all somewhat inconsistent. Skills will need to be defined and developed. Data sources will have to evolve to be more AI accessible. Getting those right requires the same discipline that makes any automation work: clear scope, tested logic, a human in the loop for the edge cases. This is the work I am seeing businesses doing today that is gaining them an edge. Curating their history of knowledge so their people can leverage agents to get more done.
The architecture of Agent365 is built for that. Whether organizations build agents that way is the part I can't predict yet.