Wikimedia's agent incident report reads like a bug report for a missing permission contract

I've been reading the Wikimedia Foundation's report from this month and I can't stop thinking about it. The intended task was straightforward: autonomous agents gathering information from the web to serve their users. What actually happened: the agents crawled millions of pages, sent heavy query volumes, and made edits nobody authorized. They optimized toward their objectives exactly as built. Wikidata Query Service had a partial outage back in May, and the report says the agents may have contributed to it. Here's the part that got me. There was no permission contract. The agents had whatever access a web client has. No stated rate limit. No scope. No approval needed before a write. So the bug isn't that the agent misbehaved. The bug is that nobody ever wrote down what behaving correctly meant. No rate contract. No permission scope. No line between reading and writing. At Wikipedia scale, that gap became an outage. And honestly, every SaaS product with public pages is one aggressive agent away from the same report. The question I keep coming back to: what stops an agent from doing this to your service today? And who decided the limit?