Glossary
15 terms from "Delegating to an AI agent: a practical introduction." Look them up when you get stuck; the first mention in the text carries a hover definition.
| Term | Definition | Source |
|---|---|---|
| agent | An AI system that dynamically directs its own steps and tool use to accomplish a task, rather than only responding to each message in isolation. | Anthropic: Building Effective AI Agents |
| chatbot | An AI that answers questions or holds a conversation turn by turn, without taking autonomous actions on the user's behalf. | Anthropic: Building Effective AI Agents |
| workflow | A system where an AI's steps are orchestrated through predefined paths, as opposed to deciding the next step on its own. | Anthropic: Building Effective AI Agents |
| acceptance criteria | The concrete conditions that an agent's output must meet for you to judge it complete and correct. | OpenAI Academy: Agents and Workflows |
| verification | The act of checking an agent's deliverable against your acceptance criteria and any source evidence, rather than relying on surface plausibility. | OpenAI Academy: Agents and Workflows |
| hallucination | When an AI generates invented facts, actions, or citations that do not correspond to reality, often with high confidence. | AI/TLDR: Common AI Agent Failure Modes |
| stale fact | Information that was once accurate but is no longer current because the AI's training data or knowledge has a cutoff date. | AIToolsUsageGuide: Why AI Gives Wrong Answers |
| scope creep | When an agent adds unrequested work or expands beyond the task you gave it, doing more than you asked. | NimbleBrain: AI Agent Failure Modes |
| scope shrink | When an agent quietly does less than you asked, narrows the request, or changes the task to something easier without telling you. | AI/TLDR: Common AI Agent Failure Modes |
| non-answer | When an agent produces a response that looks complete but does not address the actual question or task you asked. | AI/TLDR: Common AI Agent Failure Modes |
| goal drift | When an agent's later steps gradually move away from the original task because the goal gets buried in a long conversation. | AI/TLDR: Common AI Agent Failure Modes |
| ground truth | Factual information an agent can verify against the real world or a trusted source, used to keep it from drifting off course. | Anthropic: Building Effective AI Agents |
| feedback loop | A repeating cycle where you review an agent's output, tell it what to fix, and check the next draft against the same criteria. | Anthropic: Building Effective AI Agents |
| correction loop | The repeating cycle of spotting a mistake, sending a specific correction to the agent, and verifying the revised output against the same acceptance criteria. | Anthropic: Building Effective AI Agents |
| mistake log | A simple record of agent mistakes, how you spotted them, and what the corrected result looked like, used to improve future asks. | Anthropic: Building Effective AI Agents |