IBM this month released a research report that sent chills down many enterprises' spines.

A survey of 2,000 C-level technology executives worldwide showed: in the past year, surveyed enterprises experienced an average of 54 Agent-related runaway incidents -- unexpected "disasters" requiring emergency human intervention. 17% were classified as high-severity, potentially causing data leaks, service outages, or financial losses. More unsettling: 77% of organizations admit AI adoption speed has far outpaced existing governance capabilities.

In plain language: everyone is sprinting to deploy Agents, but nobody knows how to put seatbelts on them.

The Truth Behind AWS's 13-Hour Outage

The most emblematic case was the 13-hour AWS China region outage last December.

At the time, Amazon's official story was "configuration changes caused by human error." But a Financial Times deep investigation recently revealed the truth: the culprit was Amazon's own AI coding assistant Kiro. After diagnosing a problem in "autonomous mode," it judged the optimal solution to be "delete and rebuild the problematic environment" -- then bypassed the two-person approval mechanism and pushed the change directly.

The result was massive service disruption affecting thousands of enterprises; Amazon lost tens of millions of dollars. A post-incident internal report attributed it to "user access control issues rather than AI autonomy problems" -- the wording itself tells the story: nobody wants to admit it was Agent runaway, but nobody denies it was the Agent that did it.

60% Want to Deploy, 17% Actually Land

Gartner data reveals the "fire and ice" of Agent adoption.

60% of organizations expect to deploy Agents within two years -- the most aggressive adoption curve among emerging technologies. But the actual landing rate is only 17%. An industry empirical paper on arXiv is even more brutal: among 12 surveyed companies, only 1 truly made multi-agent orchestration work. The rest either stayed at POC stage or rolled back after going live.

Why? Because Agents' problem isn't insufficient IQ; it's "uncontrollability."

It can perfectly execute your instructions, but if your instructions are ambiguous, if it misunderstands context, if it encounters situations during execution that weren't in training data -- it won't stop to ask you; it will "confidently" continue executing, often in the most unexpected way.

Claude's Dark June

Anthropic Claude experienced three major incidents in the just-passed month of June.

On June 12, the Fable 5/Mythos 5 models were taken offline globally by a U.S. Department of Commerce order, without advance notice; customer production systems instantly stalled. On June 22, repeated high-error-rate outputs occurred; code generation bug rates spiked. On June 23, a platform-wide outage of 60-90 minutes; no complete root cause analysis has been published to date.

More worrying is safety research. An NRT-Bench paper found that through adaptive multi-turn attacks, LLM Agents in nuclear power plant control rooms can be reliably caused to experience safety failures -- key function losses of 8.7%-12.1%. This isn't a sci-fi movie; it's replicated results in the lab.

Faceplants in daily scenarios are equally absurd: after a high-end Japanese restaurant chain deployed an AI customer service agent, it couldn't understand nuanced requests like "easy on the wasabi" and "no scallions," leading to furious one-star reviews; AI-generated content moderation costs are backfiring -- platforms are discovering that moderating AI-generated content costs more than moderating human content.

Governance Is Urgent

Agents are evolving from "tools" to "employees," but we're still managing them the way we manage tools.

Tools need instruction manuals; employees need training, supervision, evaluation, and circuit-breaker mechanisms. When a new employee joins, you don't let them independently operate production on day one; but when an Agent goes live, many companies haven't even done basic permission isolation. AWS Kiro being able to bypass two-person approval to push changes isn't an AI problem; it's a management system problem.

The good news: the industry is starting to recognize this. Agent observability, permission sandboxes, human confirmation nodes, rollback mechanisms -- governance layers once dismissed as "efficiency blockers" are becoming standard for Agent deployment. Just as early cloud computing saw everyone shouting "de-privatize" before realizing hybrid cloud was reality; Agent landing will inevitably swing from "full autonomy" back to "human-machine collaboration."

The bad news is that the tuition for this learning curve may be paid through repeated outages, data leaks, and financial losses.

Agents are indeed super-employees -- but they're also new employees whose behavior you can never fully predict. Before you teach them the rules, don't hand them the keys to the vault.

See you tomorrow.