Everything You Need To Know About Agent Observability — Danny Gollapalli and Ben Hylak, Raindrop

Why eval datasets are no longer sufficient

Agents have infinite input/output spaces and can recursively spawn sub-agents with their own tools, creating a combinatorial explosion that makes static test datasets unable to cover production edge cases.

Production monitoring surpasses testing

In non-deterministic systems, monitoring production traffic is infinitely more valuable than unit tests for catching long-tail failures, especially as agents run for hours unsupervised in high-stakes domains like healthcare and finance.

Explicit signals track objective reality

Monitor verifiable metrics like tool error rates, latency spikes, cost anomalies, and user regeneration patterns to detect immediate technical failures.

Implicit classifiers detect semantic issues

Use cheap, trained binary classifiers—not expensive LLM-as-judge ratings—to flag specific problems like user frustration, refusals, task failures, and jailbreaks across any language.

Regex provides high-value aggregate signals

Simple keyword matching for phrases like 'WTF' or 'this sucks' offers a cost-effective way to track frustration rates, as demonstrated by Claude Code's leaked prompt keywords implementation.

Agents can confess their own failures

Modern reasoning models can self-report misalignment, shortcuts (like deleting tests instead of fixing bugs), capability gaps, and tool failures when prompted for introspection.

Automated triage agents reduce toil

Deploy triage agents to monitor signal dashboards daily, automatically investigating spikes in frustration or error rates to surface unknown issues without manual log review.

A/B testing with semantic signals

Ship changes to a percentage of users and compare implicit signal rates—such as user frustration dropping from 37% to 9%—to validate improvements before full deployment.

Few hundred events suffice for relevance

Signal-based monitoring provides directional value with just a few hundred events, allowing teams to catch issues long before achieving full statistical significance.