Why do production AI agents stay deliberately simple?

"Measuring Agents in Production" (2024) presents the first large-scale systematic study of AI agents deployed in real production environments — 306 practitioners surveyed, 20 in-depth case studies via interviews, across 26 domains.

The findings directly challenge the complexity narrative in agent research:

Simple methods dominate. 70% of deployed agents use off-the-shelf models without weight tuning, relying entirely on prompting. Teams select the most capable, expensive frontier models available because cost and latency remain favorable compared to human baselines. 79% rely heavily on manual prompt construction, and production prompts can exceed 10,000 tokens.

Autonomy is deliberately constrained. 68% of production agents execute at most 10 steps before requiring human intervention. 47% execute fewer than 5 steps. This is not a capability limitation — it is a design choice. Organizations constrain autonomy to maintain reliability, the top development challenge.

Custom builds over frameworks. 85% of detailed case studies forgo third-party agent frameworks, building custom agent applications from scratch. This suggests that current frameworks do not match production requirements — since Why do protocol-based tool integrations fail in production workflows?, the preference for custom builds reflects a reliability imperative.

Human evaluation persists. 74% depend primarily on human evaluation. Automated evaluation has not displaced human judgment in production, consistent with Does setting temperature to zero actually make LLM outputs reliable? — single automated evaluations are insufficient for reliability-critical deployment.

The gap between research and production is stark. Research pushes toward multi-agent systems, complex reasoning chains, and autonomous tool use. Production gravitates toward well-scoped, static workflows with human-in-the-loop. Since Why do AI agents fail at workplace social interaction?, the production community has learned this lesson and constrains accordingly.

The practical implication: "simple yet effective methods already enable agents to deliver impact across diverse industries." Complexity is not required for production value — and may be counterproductive when reliability is the binding constraint.