Microsoft’s explanation for its November 12 outage is worth reading closely. In July, Microsoft shipped a fix to the gateway servers that sit in front of much of Microsoft 365. The fix had a latent lock contention bug, worse on newer, higher-core hardware than the machines it was load tested on. It ran clean for more than three months. Then a routine Teams failover moved traffic onto the gateway Exchange Online also uses, on a day when about 14 percent of gateway capacity was already out for an unrelated reason. Servers stopped responding, traffic management shifted load onto fewer and fewer of them, and within two hours more than 80 percent of the fleet was unresponsive. (ITPro has the details.)
Nothing in that chain was reckless. Each piece was fine on its own. The outage lived in how they connected.
It was that kind of week. The FAA stopped flights nationwide after a router install. A Workday disaster-recovery test reached production databases. Apple traced its outage to an internal update. We keep a running list on our Outage Watch page.
Then on Saturday, Yahoo Finance ran a piece asking whether AI coding tools could make all of this worse. The argument: AI writes code faster than people can review it, so nobody really learns how the system works, and when it breaks nobody knows what else depends on it.
The article builds on an essay by Addy Osmani, who works on Claude Code at Anthropic. He calls it comprehension debt: the gap between how much code you have and how much of it anyone actually understands. I think he’s exactly right. For those of us who run infrastructure, I’d push it one step further.
Review was the safety net
Code review has always been where understanding came from. Someone reads your change, asks questions, and walks away knowing the system a little better.
It doesn’t work when an agent can open twenty pull requests before lunch. Tests, AI reviewers and a quicker human pass will pick up some of the slack, but each of those also takes out a person who used to have to understand what was going on. And the understanding really does slip. In a study Anthropic published in January, developers who used AI help while learning a new library scored 17 percent lower on a quiz about code they had just written. The biggest gap was in debugging, which is exactly the skill you need at 2 a.m. when production is down.
Most of what breaks isn’t in the diff
Even if you could review every line, you’d miss most of what takes you down, because the dependencies that matter in an outage usually live in the running system, not the repo:
- An IAM role whose trust policy lets another account’s pipeline assume it
- A security group that references another team’s security group instead of a CIDR
- A private hosted zone or Route 53 alias three services resolve through
- A KMS key policy that quietly gates every read of a shared bucket
- A NAT gateway or Transit Gateway route table that half the VPCs egress through
- The resource somebody created in the console last year that your Terraform state has never heard of
None of that shows up in a pull request, and thanks to drift, plenty of it isn’t in your IaC either.
If you’ve ever owned an old house, you know this problem. The breaker labeled “kitchen” also runs the freezer in the garage. Nobody wrote that down, and you find out when the ice cream melts. AWS’s October 2025 us-east-1 outage was a cloud-scale version: a latent race condition in DynamoDB’s DNS automation left an empty record for the regional endpoint, and everything that resolved through it, including other AWS services, went with it.
Make the system easy to ask about
If reading more code can’t close the gap, something else has to. Before anyone changes something, a person or an agent, they should be able to answer a few questions in seconds, from live state and not from a diagram someone drew last year:
- What depends on this, directly and transitively?
- Who owns it?
- What can’t be undone?
- If it goes wrong, how do we recover?
That goes double for the fix you make in the middle of an outage. The pressure is highest right when nobody has time to think through side effects, and as the article points out, a fix made during an outage can cause the next one.
The Anthropic study had a hopeful finding too. Developers who used AI to ask questions and get explanations kept their understanding. The same is true in operations. If your AI can explain how your systems connect, it builds understanding instead of draining it.
This is the problem we started OpsCanvas to solve. We map your cloud into a live Cloud Intelligence Graph™ of what runs, what depends on what, and who owns it, and plug it into the AI tool your team already uses. Anyone on the team, or their AI, can check the blast radius of a change before it ships. On Pro, the whole team works from the same map, so what one engineer knows doesn’t stay in one engineer’s head. Building the graph uses only read-only actions, and a person approves any change.
”The AI wrote it” won’t cut it
The article ends on accountability, and I think that’s where this is really going. When a serious outage hits a bank, a hospital or an airline, “the code was AI-generated” won’t excuse anything in the post-incident report. Neither will “we didn’t know that depended on it.”
I’m bullish on AI. It’s going to keep making change faster, and that’s mostly a good thing. But our ability to understand our systems has to speed up too. Otherwise we’ll keep finding out what depended on what the same way we did last week: from our customers.
Key Takeaways
Key points
- ✗Comprehension debt is the gap between how much code a team ships and how much of it anyone understands. AI coding tools are widening it fast.
- ✗More review won't close it. Nobody can read code as fast as AI writes it, and many of the dependencies that cause outages aren't in the code anyway.
- ✗Last week's outages fit the pattern. Microsoft's started with a July fix that ran clean for three months, until a routine Teams failover and 14 percent of capacity already offline pushed it over the edge.
- ✓What helps is a live map of what depends on what, so anyone can check what a change touches before it ships.
- ✓How you use AI matters. Developers who used it to ask questions kept their understanding. Use AI to explain your systems, not just to change them.