How to Find the Agent Failures Your Evals Miss [Scott Clark] - 767

2026-05-07 · Source: The TWIML AI Podcast with Sam Charrington · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics, Robotics & Autonomous Systems · Depth: Advanced, extended

Summary

Scott Clark, co-founder and CEO of Distributional, introduces "Maslov's hierarchy of observability" for AI systems, comprising telemetry, monitoring, and analytics. Telemetry, the base layer, involves logging system activity for debugging. Monitoring, the next layer, focuses on real-time detection of known signals like response times or profanity. The top layer, analytics, uses unsupervised learning to discover "unknown unknowns" or anti-patterns in production data, enabling iterative self-improvement and self-healing of AI agents. Distributional's tool, available for free on-premise or as SaaS, enriches traces into vectors, clusters them to find sub-distributions, and uses LLMs to explain differences and suggest fixes. This approach helps identify issues like agent "laziness" or "cheating" in tool calls, which traditional monitoring might miss, especially in non-stationary environments where underlying models frequently shift.

Key takeaway

For AI Architects deploying agentic systems, you should prioritize a layered observability strategy. Start with comprehensive OpenTelemetry logging using the GenAI semantic convention, then implement real-time monitoring for known signals. Crucially, integrate post-production analytics to automatically uncover "unknown unknowns" and anti-patterns that impact reliability and trustworthiness, allowing for continuous, data-driven refinement of your agents in dynamic production environments.

Key insights

Analytics, atop telemetry and monitoring, uncovers unknown AI system anti-patterns for continuous improvement.

Principles

• AI system trust requires understanding, not just benchmark optimization.
• Non-stationarity necessitates online learning and adaptive analytics.
• Effective evals demand recursive refinement and task-specific metrics.

Method

Map traces to vectors, cluster high-dimensional distributions to find sub-patterns, use LLMs to explain differences, and suggest fixes for iterative system refinement.

In practice

• Instrument with OpenTelemetry and GenAI semantic convention.
• Use analytics to discover new metrics for monitoring.
• Feed analytical insights into fine-tuning or synthetic data generation.

Topics

• Maslov's Hierarchy of Observability
• AI Agent Analytics
• LLM Anti-Patterns
• Non-Stationary Models
• Data Flywheel

Best for: AI Architect, MLOps Engineer, AI Engineer, Director of AI/ML

Related on AIssential

• Your Agent Gave the Right Answer for the Wrong Reason — and You Have No Idea — Agentic AI systems require deep observability and multi-dimensional evaluation to prevent silent failures and ensure reliable, efficient operation.
• How to Find the Agent Failures Your Evals Miss with Scott Clark - #767 — Effective LLM operations require a hierarchy of observability, with analytics crucial for discovering unknown issues and driving iterative improvement.
• Designing AI-Driven Observability for Trustworthy Agentic AI Systems — Agentic AI systems require AI-native observability to ensure trustworthiness, moving beyond traditional infrastructure monitoring.
• The Invisible Failure Mode of Agentic AI — Agentic AI failures are often invisible resource issues, not reasoning flaws, requiring real-time, local observability.
• Stop Shipping on Vibes — How to Build Real Evals for Coding Agents — Robust AI evaluations are crucial for quantifiable shipping decisions, performance monitoring, and iterative improvement.
• Whitepaper Companion Podcast - Agent Quality — Agent quality requires architectural integration, focusing on the entire execution trajectory, and continuous evaluation with robust observability.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by The TWIML AI Podcast with Sam Charrington.