AI agents glossary
Key terms from this course, in plain English.
- Agent
- A system where an LLM dynamically directs its own process and tool use, typically in a loop with environmental feedback.
- Agent loop
- The cycle of deciding, acting with a tool, observing the result and repeating until done or a limit is reached.
- Augmented LLM
- An LLM enhanced with tools, retrieval and memory — the basic building block of agentic systems.
- Escalation
- Handing a task to a human, with a summary, when an agent should not proceed alone.
- Evaluator–optimiser
- A pattern where one LLM generates and another evaluates and gives feedback, iteratively.
- Function calling
- Another name for tool use: the model outputs a structured request to call a function.
- Guardrail
- A control that constrains what an AI system can do, such as permissions, validation or approval.
- Human-in-the-loop
- Requiring human approval or input at defined points in an AI process.
- Idempotent
- An action that has the same effect whether performed once or several times.
- Long-term memory
- Information an agent stores and retrieves across sessions, e.g. in a database.
- Model Context Protocol (MCP)
- An open standard for connecting AI applications to servers that expose tools, resources and prompts.
- Orchestrator–workers
- A pattern where a central LLM dynamically breaks down a task and delegates to worker LLMs.
- Parallelisation
- Running LLM calls simultaneously — sectioning into subtasks or voting across attempts.
- Prompt chaining
- A workflow of sequential LLM calls, each processing the previous output.
- Routing
- Classifying an input and directing it to a specialised prompt, tool or model.
- Shadow mode
- Running an agent to propose actions without executing them, to compare with human decisions.
- Skill
- A packaged, reusable set of instructions (and optionally scripts) an agent can load when relevant.
- Sub-agent
- An agent invoked by another agent to handle a sub-task with its own instructions and context.
- Tool
- A function an LLM can request, described by a name, description and input schema.
- Trace
- A record of an agent run: prompts, tool calls, results, decisions, timing and cost.
- Trajectory
- The sequence of steps and tool calls an agent takes to complete a task.
- Workflow
- A system where LLMs and tools are orchestrated through predefined code paths.