read

Estimated Reading Time: 4 minutes

Software used to change at the speed people could type. Today AI agents write code, open pull requests, ship configuration and increasingly operate the systems they helped build. The volume of change has multiplied, while the number of humans who truly understand any given system has not.

That gap is where outages live. AIOps and AI SRE, once a nice efficiency story, are now the only practical way to keep pace.

The evidence is already in the headlines

Recent months have shown how fragile large systems already are. On 12 November 2026 a large number of Exchange Online servers failed health checks, cutting users off from Outlook, Teams and Copilot, and restored servers fell over again once traffic returned. A day later an authentication failure locked people out of Claude, while its Microsoft 365 connector was already failing, a reminder that one outage now travels down the dependency chain.

Airlines tell the same story. An IT failure at American Airlines led to a nationwide ground stop in July 2026, and that same month a fault in the United Airlines reservation system disrupted check in and boarding at airports across the United States. Before that, a Cloudflare configuration error in November 2025 and a 15 hour AWS outage in October 2025 each took large parts of the internet down with them.

None of these has been blamed on AI agents. That is the point, our systems are this fragile with humans in charge, and agents are about to add far more change on top.

The operational maths no longer works

Classic SRE assumed a human could hold the system in their head. An on call engineer read the dashboards, formed a hypothesis and followed a runbook. That model strains when thousands of changes land every day, many authored by agents, across services no single person has read.

The systems themselves have changed too. An agentic application is non deterministic, it can take a different path on every request, call tools in a new order and fail in ways no alert was written for. Static thresholds and hand written runbooks cannot describe that behaviour.

AI SRE closes the gap by putting machine reasoning where the machine scale problem is. It correlates signals across metrics, logs, traces and changes, proposes a root cause in minutes and carries out safe remediation, leaving humans to judge the hard calls.

Reliability becomes the product

When an agent acts on behalf of a customer, an outage is no longer a slow page. It is a booking not made, a payment not reconciled, a case left unanswered. Reliability stops being an infrastructure metric and becomes the thing the customer is buying.

Agents also raise the blast radius of mistakes. A human misconfigures one service, an agent with broad permissions can repeat the same error across a fleet before anyone notices. So the agents that run production need what we ask of any senior engineer:

  • Evidence before action, every conclusion traceable to the signals behind it
  • Bounded authority, with staged rollouts and automatic rollback
  • Verification after the fix, confirming that the system actually recovered
  • A full record of what was done and why

Security moves to the front of the queue

Every agent is a new identity with credentials, tool access and the ability to act. That makes it a target. Prompt injection, poisoned context and overly broad permissions turn a helpful assistant into an insider threat that works at machine speed.

An AI SRE agent is the sharpest case, because it reads telemetry from everywhere and can change production. It must run with least privilege, treat every log line and ticket as untrusted input, and require approval for actions that cannot be undone.

Security and reliability are now the same discipline. An incident caused by an attacker and an incident caused by a bad deploy look identical at three in the morning.

Why the air gap matters again

For years the air gap looked like a relic from before the cloud. Agents have brought it back. Operational data is among the most sensitive an organisation holds, it reveals architecture, customer behaviour, secrets that leak into logs and the exact shape of every weakness.

Sending that stream to a model outside the boundary is a decision many regulated industries, governments and critical infrastructure operators cannot make. They need the intelligence to come to the data, with the models, the agent harness and the knowledge all running inside the perimeter, with no call home.

There is a resilience argument too. The system that diagnoses your outage should not depend on a network path or third party service that may be part of the outage. Capable open weight models now make this practical, so an air gapped AI SRE is no longer a compromise, it is a design choice.

What to do now

Treat AI SRE as core infrastructure, not a tooling experiment. Give your operational agents the same rigour you give production code: evaluation against real incidents, tight permissions, complete audit trails and a deployment model that fits your data obligations, including fully isolated environments.

The organisations that win the agent era will not be those that ship the most agents. They will be the ones whose agents, and the systems beneath them, can be trusted to stay up and stay safe.

The views expressed here are my own.

Sources

Blog Logo

Manpreet & Renaira


Published

Image

The Tokens

Bonding through words, and the math behind them.
Manpreet & Renaira

Back to Overview