Measuring Agents in Production

TL;DR

This paper provides a systematic study of how large language model-based agents are deployed in production, revealing common practices, challenges, and the reliance on simple, controllable methods based on practitioner data.

Contribution

It is the first comprehensive analysis of production deployment practices for LLM agents, based on interviews and surveys with practitioners across multiple domains.

Findings

Most agents execute fewer than 10 steps before human intervention.

Majority rely on prompting off-the-shelf models rather than weight tuning.

Reliability remains the top challenge addressed mainly through system design.

Abstract

LLM-based agents already operate in production across many industries, yet we lack an understanding of what technical methods make deployments successful. We present the first systematic study of Measuring Agents in Production, MAP, using first-hand data from agent developers. We conducted 20 case studies via in-depth interviews and surveyed 306 practitioners across 26 domains. We investigate why organizations build agents, how they build them, how they evaluate them, and their top development challenges. Our study finds that production agents are built using simple, controllable approaches: 68% execute at most 10 steps before human intervention, 70% rely on prompting off-the-shelf models instead of weight tuning, and 74% depend primarily on human evaluation. Reliability (consistent correct behavior over time) remains the top development challenge, which practitioners currently address through systems-level design. MAP documents the current state of production agents, providing the research community with visibility into deployment realities and under-explored research avenues.