ICA · Institutional Control Architecture
An essay from ALEETH  ·  November 2026
ALEETH · Essay

The Outage Test

On November 12, Microsoft 365 had a major outage. When AI agents and the controls on them share one cloud, one outage takes both. ICA governs from outside, and holds what matters when it cannot see.


On Thursday, November 12, 2026, Microsoft told its customers they might be unable to connect to Exchange Online, Microsoft Teams, SharePoint Online, Microsoft 365 Copilot and Microsoft Defender XDR. The incident began at 1:26 PM Eastern. By evening Microsoft was reporting only limited recovery. Every way of governing AI agents eventually takes the same test: what happens on the day the platform does not answer.

The Outage

Work, AI and security, on one outage list

Microsoft first attributed the disruption to elevated load across several regions, then found the cause: a number of servers being unexpectedly unavailable; straining the remaining infrastructure. It paused all deployments and was moving traffic away from the affected systems and restarting servers to restore capacity. The advisory, tracked as MO1550821, said users in any region could be affected. Earlier the same day, a separate Teams incident had disrupted messaging across North and South America for just over an hour.

Look at the list again. The mail, the chat, the AI assistant, and the security platform meant to watch for trouble. For some customers, the tool that should notice something going wrong was one of the services they might not have been able to reach.

The Scale

Agents at the size of the workforce

The industry is already planning for agents in enormous numbers. In a June interview that went viral this month, Microsoft's chief executive, Satya Nadella, pictured a company of 200,000 employees with, in his words, let's say 2 million agents or 20 million agents all in a loop. He also said such agents need to be fully inspectable, fully auditable.

The clip spread less than two weeks after Microsoft offered voluntary exits to about 60,000 employees. Microsoft says AI is not behind that offer and expects fewer than 5 percent to take it. The question it raises stands anyway: when the agents number in the millions, where do the controls on them live?

A control that goes down with the system it controls is a convenience, not a control.
The distinction
The Gap

One cloud, one point of failure

When an agent's model, its tools, its identity, its security monitoring and its audit trail all run on one vendor's platform, the outage that stops the work can also blind the oversight. The institution is left to trust that whatever happened during those hours was fine, with the record of it sitting inside the system that failed.

That is not a criticism of any one vendor. Every platform has a bad day. It is a question of architecture: whether the authority over an agent sits inside the agent's platform, or outside it.

The Answer

Govern from outside. Hold when blind.

ICA sits between an AI agent and the systems it would change, on AWS, separate from the agent's model vendor and from the productivity cloud. Every governed action stops at ICA first.

01IndependentThe control is not part of the platform it controls. An outage in the agent's vendor platform does not, by itself, take ICA's decisions or ICA's record with it.
02Fails closedFor an organization with enforcement armed, when ICA cannot see its own backend, an irreversible action is held, not passed through, and a simple read still runs. ICA's proof suite simulates a total outage and checks exactly this.
03Deny by defaultNo signed scope contract, no consequential action: anything beyond a read is refused. The gateway's decision logic also has a formal specification, which a model checker proves over every combination of request properties it models.
04SealedEvery decision becomes a signed receipt, anchored to a public log, with an RFC 3161 trusted timestamp over the log's root.
05CheckableAnyone can verify a receipt offline against ICA's published key, with no ALEETH server in the loop. A receipt you already hold still verifies during an outage of ours.
What ICA does with each governed action

Failing closed has a cost, and it should be said plainly. When ICA cannot decide, consequential work waits. That is the trade an institution should choose on purpose: an agent that pauses is a delay; an agent that acts unwatched is an exposure.

The Record

Check it yourself

ICA publishes its own enforcement coverage, gaps included. On November 13, 2026 it read 1,513 of 1,513 action paths that change anything, governed: 1,419 routes, 81 background jobs and 13 queue consumers. Its latest anchored log root carried a trusted timestamp whose signature and issuer chain checked out. Both are on the live record at aleeth.com/trust.

What It Means

The authority belongs to the institution

Microsoft builds a platform many agents will run on, and nothing here argues otherwise. The point is where the authority sits. The institution sets the limits, names the people who approve, and holds the record, outside the blast radius of any one vendor's bad day, including ours.

Twenty million agents will be inspectable and auditable only if the inspection and the audit still work when the platform does not.

Model reasons. Tools act. ICA controls, and the institution holds ICA.
The answer

Sources: BleepingComputer, "Massive Microsoft 365 outage disrupts Outlook, Teams, Copilot worldwide," November 12, 2026, with Microsoft's advisory MO1550821, its updates on X, and the earlier Teams incident TM1550488. Windows Latest, "Microsoft's CEO imagines 20 million AI agents doing the work, as the company offers 60,000 employees an exit," November 9, 2026, quoting Satya Nadella on the Possible podcast (June 5, 2026) and Microsoft's statements on the voluntary exit offer. ICA proving harness 01 (fail-closed matrix) and gateway decision specification (TLA+), commit 4877e221, read November 13, 2026. ICA public enforcement coverage and transparency status, read November 13, 2026. Independent analysis; no affiliation with Microsoft.