Glossary
8/5/2026, 5:11:49 PM
Practical definitions for the 20 core terms used throughout the Agentic AI series (agentic workflow, autonomy, tool, trace, guardrail, MCP, and more).
Glossary
| Term | Practical meaning |
|---|---|
| Agentic workflow | An LLM-based application that executes multiple steps toward an outcome |
| Autonomy | The degree to which the model chooses actions, tools, order, or stopping |
| Component contract | Typed inputs, outputs, errors, and side-effect rules for one step |
| Evaluation (eval) | A repeatable procedure for measuring system behavior |
| Error analysis | Trace-based classification of failures to identify what to fix |
| Reflection | A critique-and-revision loop applied to an initial output |
| Tool | A function the model may request the application to execute |
| Tool schema | Structured description of a tool and its permitted arguments |
| Observation | A tool or environment result returned to the agent state |
| Plan | An explicit or implicit sequence of proposed actions and dependencies |
| Orchestrator | Application logic that routes state, model calls, tools, and controls |
| Trace | Chronological record of model calls, decisions, tools, results, and errors |
| Guardrail | A technical or procedural control that constrains behavior |
| Human in the loop | A person reviews, supplies information, or authorizes an action |
| MCP | A protocol for exposing tools, resources, and related capabilities to AI clients |
| Idempotency | Repeating an operation has no additional unintended effect |
| Regression test | A retained test ensuring a previously fixed behavior remains fixed |
| Context | Information supplied to the model for its current call |
| Handoff artifact | Structured output passed from one agent or component to another |
| Termination rule | A condition that ends a loop on success, budget exhaustion, or escalation |
Learning map
Page 39 of 40 in DeepLearningAI > Agentic AI. Read after "5.7 Communication Patterns for Multi-Agent Systems". Continue to "Capstone: Build an Evaluated Research Agent" next. All 37 numbered lesson pages (1.1-5.7) share one template — state, contracts, controls, observability, and evaluation, introduced in full in 1.1 Course Overview — so this page assumes that shape and focuses on what's unique to its own topic.
Get hands-on — step by step
Keep this page open while reading the rest of the series; when a term's practical meaning is unclear in context, check it here rather than guessing.
Top 3 sources
- 1Claude Docs: Prompt Engineering Overview
Anthropic's official guidance on structuring prompts and multi-step model interactions.
https://docs.claude.com/en/docs/build-with-claude/prompt-engineering/overview
- 2DeepLearning.AI Course Catalog
The broader DeepLearning.AI curriculum this study guide's structure is organized around.
https://www.deeplearning.ai/courses/
Links are AI-suggested — worth a quick sanity check before diving in.