Meta's year-long data shows AI agents can't replace employees: incidents up 40%, engineer firefighting up 70%

InfoQ-CN reports that Meta's year-long experience with AI agents points in the opposite direction of the "agents replace employees" narrative: incidents rose 40 percent, and engineer firefighting work increased 70 percent.

Over the past year, major tech companies have raced to bring AI agents into internal workflows, and Meta has been among the most active, tasking agents with coding, operations, and data processing. The new figures suggest the hidden costs of automation — monitoring, troubleshooting, rework — may be eating into some of the efficiency gains.

A 40 percent rise in incidents implies agent adoption is not painless: when automated actions go wrong, the blast radius tends to be larger and harder to trace. A 70 percent jump in engineer firefighting means headcount that could have shipped new features is being consumed by problems agents create.

The results stand in sharp contrast to the market narrative from OpenAI, Anthropic, and others that agents are about to take over large swaths of the workforce. Over the past months, vendors have launched computer-use agents and coding agents claiming to handle parts of junior roles, and Meta's frontline data now offers a counterexample from inside a major lab.

None of this means agents lack value — the more reasonable reading is that current agents work best as assistive tools embedded in existing processes rather than as direct replacements for people. Deployment design, human-machine division of labor, and monitoring systems may matter more than raw model capability.

What to watch next: whether Meta adjusts its internal deployment strategy and repositions agents from "replacers" to "collaborators," and whether these numbers temper other companies' expectations about how quickly agents can be trusted on their own.