Glossary / 24 terms, zero hype
The agent jargon, decoded.
Every term a vendor will use at you, in two sentences each: what it is, and why it matters to your business. If a definition here does not survive contact with a real meeting, tell us and we will fix it.
24 of 24 terms shown
- Core ideasAI agent
- Software built around a language model that plans and carries out multi-step tasks towards a goal, using tools, instead of just answering one question.
- Why it matters The difference between advice and work getting done: you delegate an outcome, not a prompt.
- Core ideasLarge language model (LLM)
- The underlying AI system trained on large collections of data that can process patterns and generate language.
- Why it matters It is the engine inside many agents; the agent adds tools, state and an execution loop around it.
- Core ideasPrompt
- The instructions and context given to a model or agent: its role, task, constraints and examples.
- Why it matters Prompt quality affects the result, alongside the model, data, tools, evaluations and system design.
- Core ideasContext window
- The amount of text a model can consider at once: its working memory for a single run.
- Why it matters Explains why agents summarise and retrieve instead of reading everything every time.
- Core ideasAgentic workflow
- A pipeline of trigger, steps and output that an agent executes: read, decide, act, report.
- Why it matters Thinking in workflows turns a vague AI idea into something you can actually build and audit.
- Core ideasMulti-agent system
- Several specialised agents cooperating on one job: one plans, one researches, one writes, one checks.
- Why it matters Sounds futuristic, is often just sensible division of labour; also multiplies the governance surface.
- How agents are builtTool (function calling)
- A capability an agent may invoke: search a database, send an email, read a file, call an API.
- Why it matters The tool list is part of the boundary. Credentials, scopes, policy enforcement and runtime controls determine what it can really touch.
- How agents are builtMemory
- Stored context that survives between runs: past cases, preferences, decisions.
- Why it matters Without memory every run starts from zero; with it come privacy duties over what is retained.
- How agents are builtOrchestration
- The layer that sequences steps, routes work between tools or agents, and handles failures and retries.
- Why it matters Orchestration, not the model, decides how reliable the whole pipeline is.
- How agents are builtRAG (retrieval-augmented generation)
- The system retrieves relevant passages from a selected knowledge source before generating an answer.
- Why it matters It can ground an answer in your material, but retrieval quality and model errors still need evaluation.
- How agents are builtEmbeddings
- Numeric fingerprints of text that let software find passages by meaning rather than keywords.
- Why it matters The indexing trick behind RAG and semantic search over your documents.
- How agents are builtMCP (Model Context Protocol)
- An open protocol for connecting AI applications to tools and data sources through a consistent interface.
- Why it matters It can reduce custom interface work, while compatibility, authentication, consent and security still require engineering.
- Safety and governanceGuardrails
- Hard limits on what an agent may do: allowed tools, spending caps, blocked topics, mandatory approvals.
- Why it matters Guardrails are what make delegation to software defensible in front of a client or a regulator.
- Safety and governanceHuman in the loop
- A pipeline design where a person must review or approve before consequential actions execute.
- Why it matters It is one useful control when the reviewer has enough context, authority and time to make a real decision.
- Safety and governanceLeast privilege
- Giving an agent only the narrowest access it needs: one mailbox, not the mail server.
- Why it matters Limits the blast radius of both mistakes and misuse.
- Safety and governanceHallucination
- A model stating something false with full confidence, including invented facts, numbers or citations.
- Why it matters The reason outputs that leave the building need grounding (RAG) plus human review.
- Safety and governancePrompt injection
- A malicious input (email, web page, document) that tries to override an agent instructions.
- Why it matters Agents that read untrusted content must treat it as data, never as commands; ask your vendor how.
- Safety and governanceAudit log
- A tamper-evident record of every action an agent took, with inputs and outputs.
- Why it matters When something goes wrong, the log is the difference between a fix and a mystery.
- Running agentsEvaluation (evals)
- Structured testing of agent output quality against known-good answers, before and during production.
- Why it matters The professional alternative to "it seemed fine when we tried it".
- Running agentsBaseline
- Measuring how the process performs today (hours, errors, delays) before the agent starts.
- Why it matters Without a baseline, ROI claims are storytelling.
- Running agentsPilot
- A contained trial of one agent on one process for a fixed period, with review at the end.
- Why it matters Cheap way to learn the real effort and value before anything mission-critical depends on it.
- Running agentsOversight cost
- The human time spent reviewing, correcting and supervising an agent output.
- Why it matters The honest ROI equation subtracts it; vendors routinely forget to.
- Running agentsNo-code agent platform
- A visual builder where workflows are assembled from blocks instead of programmed.
- Why it matters Fast for standard patterns like inbox triage; custom system integrations still need technical work.
- Running agentsAPI
- An interface software uses to exchange data or request actions from other software.
- Why it matters A suitable API can make an integration more reliable, but access controls, failure handling and supported operations still matter.
Frameworks behind the glossary
- AI Risk Management Framework National Institute of Standards and Technology, 2023
- OWASP Top 10 for Agentic Applications OWASP Foundation, 2026
- Model Context Protocol architecture Model Context Protocol, 2025
Last reviewed: 2026-07-18
Jargon decoded. Now see it in motion.
The workflow builder turns these words into a working picture: triggers, steps, outputs and the guardrails between them.