This article discusses the challenges of evaluating and monitoring LLM-based agents in production, covering offline evals, prompt engineering pitfalls, observability tools, review queues, labeling, clustering, topic classification, and cost-effective layering of human review, LLM-as-a-judge, and small classifiers.
Building a chat product or autonomous agent is different from anything that came before it. Traditional products have clear metrics: did a user take a certain action? It's in your database. For conversations, *useful* is much harder to define. Was that a good interaction? What was the user even trying to do? Without evals, you're mostly guessing. Here's the monitoring layer most teams skip. **Offline evals** You need test cases your agent must pass before a new version ships. Pass/fail may not be binary, usually you define a threshold success rate for what's acceptable. The hard part is deciding what goes in. Evals need to represent production data: not the most relevant benchmark you found online, not the handful of examples from the PRD, not synthetically generated hypotheticals. If your evals don't match what actually happens in production, you're not measuring the right thing. **Prompt engineering** Past the initial wow factor, you realize the agent isn't doing what it's supposed to. So you start prompt engineering. Over time the prompt grows to tens or even hundreds of statements, and despite explicitly telling the agent that a certain behavior matters, you still see it doing the opposite in production. Often you find out by accident. That's not good enough. **Observability tools** Most LLM observability tools feel like systems monitoring dashboards rather than tools built to catch whether your agent is following your instructions. Scorers and LLM-as-a-Judge can help, but model-based approaches have their inaccuracies. You still need humans reviewing the data. Random sampling only gets you so far. You need to prioritize what to look at. **Review queues** If hundreds of conversations ask the same question, reviewing the same thing repeatedly is a waste. You need diverse examples: embedding distance, extremes in tools used, answer length, latency, or other signals. Some issues can be auto-flagged: the agent didn't follow an explicit prompt instruction, or a groundedness checker found a claim not in the knowledge base. Surface these first. **Labelling** When you review conversations, annotate them: * Flag issues with a description of the problem and why it matters. These become test cases in your offline evals. * Note the correct behavior. Specific notes on what good looks like can be used as training data. Build a taxonomy of problems specific to your application, not generic helpfulness or toxicity, but the things that actually matter for your use case. **Getting insights at scale** * **Clustering:** group similar conversations to understand what people are talking about, then drill into specific clusters * **Topic classification:** break down by use-case so you understand how your tool is actually being used; keep the taxonomy under your control * **Scorers:** a classifier or small model that adds metadata to each conversation (response length, language used, whether code was output, etc.) **Cost** Human review is irreplaceable but expensive. LLM-as-a-Judge is cheaper but costs accumulate. Small classifiers trained on human labels handle the bulk of the data cheaply. Layer them: classifiers on everything, LLM-as-a-Judge on a subsample, humans on the most ambiguous or high-value examples. How are you keeping track of your agent sessions? Curious what techniques and stacks people are using.
An article discusses the need for Agent Harness Engineering—structured systems with tool validation, context management, guardrails, telemetry, and verification loops—to make LLM agents reliable in production, arguing that better prompts alone are insufficient.
The author questions whether current LLM evaluation tools are too focused on isolated prompts rather than full workflows and agent interactions, noting that step-by-step accuracy can mask overall behavioral drift in production.
A comprehensive guide on evaluating LLM-based agent systems, covering fundamental concepts, evaluation frameworks, and case studies from recent benchmarks.
Based on conversations with over 20 teams, the author identifies three recurring pain points when using LLMs in production: enterprise-only basics, lack of agent observability, and slow support for new models.
This paper studies a deployed LLM-as-judge system for evaluating multi-turn conversational agents and finds it catches far fewer defects than human review, revealing a structured blind-spot taxonomy and routing failures.