Agent Evals
Systematic testing of AI agents against representative tasks — measuring task completion, tool-use correctness, cost, latency, and safety across many runs. Distinct from single-turn LLM evals because agent behavior spans multiple steps and depends on tool outputs, memory, and planning quality. Considered table stakes before shipping agents to production.
Related terms
- AI Agent
An autonomous AI system that can perceive its environment, make decisions, and take actions to achieve specific goals. Agents can use tools, interact with APIs, and coordinate with other agents.
- Model Monitoring
Continuous tracking of ML model performance in production, including accuracy metrics, data drift detection, and operational health. Essential for maintaining model quality over time.
- Runtime Guardrails
Enforcement layers that run around a model or agent at inference time — filtering unsafe inputs, blocking prohibited outputs, redacting PII, detecting prompt injection, and enforcing schemas. Distinct from AI governance policy: guardrails are the mechanism that makes policy actually stick during a live request.
Related on this site
Framework dimensions
Free tools
Whitepapers
Comparisons
Next steps
- Take the assessment— See where you stand
Call this framework and its tools from your own agent via the Model Context Protocol (MCP) server. Works with Claude Desktop, Cursor, Zed, Continue, and the OpenAI Agents SDK.
Where this fits
Agent Evals is part of the Agentic AI vocabulary used in the Generative AI Maturity Framework. See the full glossary for the complete set of 149 defined terms, or take the free maturity assessment to see where your organisation stands.