Microsoft Entra can now disable an AI agent's identity on its own once the agent stops requesting tokens, and the agents most likely to break are the ones that only run once a quarter. Under the default policy, Entra notifies an agent's sponsors and owners after 60 days of inactivity, disables the identity at 90 days and deletes it at 455 days, according to Microsoft's Learn page for the default inactivity policy. Before you enable it, give every agent identity a sponsor who will read the email, and exclude the year-end close bot, the audit-season agent and anything else that legitimately goes quiet for three months.
Microsoft announced the policy at Ignite on November 17, 2026, in a Microsoft Entra blog post by Yina Arenas. The same post previewed discovery of 60 local agent types, Conditional Access for agents acting on a user's behalf, and an MCP firewall. Only the lifecycle automation is labelled generally available. The discovery and firewall controls are previews, and you should pilot them as previews.
What Microsoft Shipped for Agent Identities at Ignite 2026
The generally available piece is automated lifecycle management for agent identities. The other agent controls in the release are previews. In the Ignite post, Microsoft says lifecycle policies check whether each agent "has a human sponsor, is still needed, and has been active recently," and if not, "Microsoft Entra sends notifications and later disables or deletes the agent according to your instructions." Virtualization Review's Ignite roundup reports the same split: lifecycle management is GA, while local agent discovery, agent-aware Conditional Access and the MCP firewall are in preview.
The agent items in the post break down like this:
- Generally available: automated agent lifecycle management, Conditional Access for Windows 365 for Agents, and network-layer DLP through Purview and Global Secure Access.
- Preview: Agent ID for discovered local agents ("including Claude Code, OpenAI Codex, and GitHub Copilot CLI"), Conditional Access for assistive agents, dynamic permission remediation, agent-type-aware network controls and the MCP firewall.
- Agent 365 preview: Dynamic Agent Risk, a risk score built from Entra, Defender and Purview signals.
Two caveats before you plan around the GA label. The post names no SKUs and no prices. And as of November 19, 2026, the Learn pages for both the default inactivity policy and custom lifecycle policies still carry a banner saying lifecycle policies are in preview and "aren't recommended for production workloads." Ask your Microsoft account team which statement your support contract follows.
How the 60-Day Inactivity Policy Actually Works
The default policy runs on a fixed timeline, and day 60 only starts it. Per the default policy documentation, Entra notifies sponsors and owners at 60 days inactive, disables the identity at 90 days and deletes it at 455 days. "The actions and timeframes are not customizable for default policy users."
Inactivity has a narrow meaning here. In the custom policy documentation, an agent counts as inactive when "no tokens have been issued for the identity during that time." Microsoft's Zero Trust workshop guidance on inactive agents puts it more plainly: "This is identity authentication activity, not language-model token consumption." Model spend tells you nothing about whether the agent identity is active.
The default policy has three properties that matter to an identity team:
- It costs nothing to run. "No license is required to use the Agent ID default inactivity policy." The docs list Lifecycle Workflows Administrator, AI Administrator and Agent ID Administrator as the roles that can manage it.
- The only way to keep an idle agent alive is to use it or exclude it. The docs say "Users cannot freely retain inactive agents that are subject to the inactivity policy without making them active." The scope option "All Agent IDs except selected" lets you exempt named identities, which "can remain inactive indefinitely."
- The setup steps begin with an Enable policy toggle. The Learn procedure goes ID Governance, Lifecycle Workflows, Lifecycle Policies, Agent ID (Preview), then "Toggle Enable policy." Check what state your tenant is in before you assume either way.
Custom policies add more rules. Periodic reconfirmation asks a sponsor to confirm the agent is still needed every 30 to 730 days, a minimum-sponsor rule flags agents with too few sponsors, and you choose the inactivity window yourself. Actions are Disable, Delete, or Disable and delete. A custom policy needs Microsoft 365 E7, or a Microsoft Agent 365 license paired with Entra P1 or Microsoft 365 E3. The default policy is always evaluated last, after any custom policy that matches.
Where the Policy Breaks Real Workloads
The policy can't tell a dead agent from one that runs rarely, and Microsoft's own guidance says so. The Zero Trust inactive-agent guidance warns that "a gap in available telemetry is not proof that an identity is unused" and that "disabling or deleting an inactive identity can interrupt a legitimate workload that runs infrequently."
Consider which agents in your tenant go quiet for 90 days on purpose: a quarter-end reconciliation agent, an annual access-review helper, a disaster-recovery runbook agent, a tax-season intake bot. Each gets disabled on schedule unless someone excludes it or answers the notice.
The recovery path makes this worse. A sponsor is the person who gets the warning, but per Microsoft's owners, sponsors and managers page, "Sponsors can't re-enable or restore agent blueprints or identities." Only an owner or an administrator can. If your quarter-end agent is disabled at 2 a.m. on the last day of the quarter, the business person accountable for it cannot fix it. Owners are optional on every agent identity object, so for many agents an administrator is the only person who can bring it back.
The Zero Trust guidance adds one more caution for incident responders: "Disabling an identity is not proof that every existing agent session has ended." Treat a lifecycle disable as hygiene, and keep your incident playbook's token revocation step.
Why Sponsors Come Before the Switch
Every warning goes to an agent's sponsors and owners, so an agent without a reachable sponsor is disabled with nobody reading the notice. Microsoft defines a sponsor as the business person "accountable for the agent's purpose and lifecycle decisions," and requires at least one on every agent identity and blueprint at creation. Blueprint principals are exempt.
That requirement covers new objects. It doesn't help when a sponsor changes jobs or leaves, which is the common case in a fleet created by citizen developers. Microsoft's governance overview says sponsorship "can be automatically transferred to their manager" through Lifecycle Workflows, and the Agent ID what's-new page lists two sponsor templates to "prevent orphaned agents." For agents with no sponsor or owner at all, the default policy sends notices to fallback recipients you enter, and group mailboxes are supported.
Microsoft ran a version of this cleanup on its own estate before shipping it. Its customer-zero program paired risk-tiered review with a 60-day cleanup across 20,000 makers. The sponsor model also matches what we found in the Okta, Entra Agent ID and SailPoint comparison: Entra issues and governs the identity, but it can only govern agents registered as Agent IDs.
What Local Agent Discovery and the MCP Firewall Cover
The discovery and firewall previews fill a gap that OAuth logs leave, but each needs infrastructure many tenants haven't deployed. Microsoft says Agent ID for discovered local agents works through Microsoft Defender together with Global Secure Access and gives each discovered agent, such as Claude Code or Codex, an Agent ID. The post does not publish the full list of 60 types.
The network half has hard prerequisites. Global Secure Access AI agent discovery sees local agents only on endpoints running the GSA client "with internet access forwarding policy and TLS inspection enabled," and it classifies each agent as managed or shadow by whether it is registered with Entra Agent ID. The endpoint half has its own limits. Defender's local agent discovery covers Windows and macOS, counts Claude Code running in 15 project folders for one user on one device as a single agent, and states that "an inventory entry isn't proof of active use."
The MCP firewall is narrower than its name. Microsoft's MCP firewall workshop guidance says it inspects remote MCP traffic (JSON-RPC 2.0 over streamable HTTP and Server-Sent Events) and can allow or block servers and individual tools. "Local/stdio servers and JSON-RPC batches are not inspected," and it "does not replace MCP OAuth/token validation or application authorization." Coding agents commonly launch local stdio MCP servers, so the firewall governs only part of what a GitHub Copilot or Claude Code user can reach. Our piece on MCP dependency drift covers the local half, and the shadow AI discovery buyer's guide explains why OAuth logs alone miss these installs.
What to Do Before You Enable It
This Week:
- Export every agent identity in the tenant and list the ones with no sponsor, a sponsor who has left, or no owner. Those are the agents that will be disabled with nobody reading the notice.
- Ask each business unit for agents that run monthly, quarterly or yearly. Add them to the "All Agent IDs except selected" exclusion before anyone enables the default policy.
- Set a fallback recipient to a monitored group mailbox owned by the identity team.
This Month:
- Deploy the sponsor lifecycle workflow templates so sponsorship moves to a manager when someone leaves.
- Name an owner or a small admin rota for every business-critical agent, because sponsors cannot re-enable what the policy disables.
- If you hold E7 or Agent 365 licences, test a custom policy on a handful of test agents with the 30-day reconfirmation minimum, as the custom policy docs recommend, before widening scope.
Before Your Next Agent 365 or E7 Renewal:
- Pilot local agent discovery on one engineering group that already runs Defender and the GSA client with TLS inspection, and compare its count of Claude Code and Codex installs with your OAuth logs.
- Pilot the MCP firewall only for remote MCP servers, and keep your stdio allowlist and endpoint controls in place.
- Get written confirmation of the support status of lifecycle policies, since the blog and the Learn pages currently disagree.
The Bigger Picture
Agent identities are going through the same cleanup cycle service accounts went through a decade ago, only faster. Service accounts piled up for years before access reviews caught them. Microsoft is shipping automated disable for agents within months of making Agent ID generally available, and its Ignite post cites internal telemetry showing tool and agent abuse made up 12% of attack activity observed on Azure AI workloads. Cleanup policies for human accounts got tuned over years of false disables. Agent identities will need the same tuning, and an exclusion list built from what the business units tell you is the quickest place to start. Assign sponsors and build the exclusion list first, then turn on the 90-day disable.
Continue Reading
- Okta vs Entra Agent ID vs SailPoint: Two Issue, One Governs
- Your Shadow AI Inventory Needs OAuth Logs, Then Endpoint Coverage
- Microsoft says Risk-O-Meter sorts agent builds green, yellow or red
- Copilot Studio Agents Keep Shared Memory, No User Opt-Out
- Microsoft Adds Hard Spending Caps for Copilot Studio Agents
- SK Telecom's AI Governance Report Gives Agents Employee IDs
