Custom AI Agent Glossary

A system that uses a language model to carry out a task on someone's behalf: it decides its next step, uses tools to read data and take actions, and stops or hands back to a person when it should. Anthropic distinguishes agents, where the model directs its own process, from workflows, where the steps are fixed in code.^[1][2]

Gartner's term for rebranding existing assistants, robotic process automation or chatbots as agents without real agentic capability.^[3]

The measured performance of a workflow before the agent: time, cost, volume and error rate. Without one, a return can't be shown.^[3]

Software that answers questions in a conversation. The user drives every turn, and it rarely takes actions in other systems.

An AI assistant built into a product that helps a person while they work, by drafting, suggesting or summarizing. The person decides each step and finishes the work.^[4]

An agent built for one organization's workflow, working inside its systems and under its controls, as opposed to a ready-made product sold to every customer. The model is usually commercial; the instructions, tools, integrations, guardrails and tests are custom.^[1]

Testing an agent or model on a set of cases and scoring the results, automatically where answers are checkable and by expert review where they are not.^[5][1]

Giving an agent more permissions, tools or autonomy than its job needs, so a mistake or attack can do more damage. Named in OWASP's Top 10 for LLM applications.^[6]

The known right answer for a test case, used to score an agent's output objectively. Benchmarks such as BORE rank models only on exact checks against ground truth.^[5]

Checks around an agent that keep it within bounds: validating inputs, screening for unsafe or off-topic requests, limiting which tools and actions it may use, and checking outputs before they are used.^[1]

Output from a language model that sounds plausible but is false, such as an invented citation. The reason outputs that matter need verification.^[7]

A design where a person reviews, approves or can override an agent's work at set points, typically before irreversible or high-risk actions and when the agent repeatedly fails. Some regulations require human oversight for high-risk uses.^[1][8]

The written guidance that defines what an agent does and how: its steps, rules, tone and when to escalate. Good instructions are often drawn from an organization's existing procedures.^[1]

An open standard for connecting AI applications to external tools and data sources in a consistent way.^[9]

Several agents working together, each with its own instructions and tools, coordinated by a manager agent or by hand-offs between them. OpenAI recommends starting with one agent and splitting only when it becomes hard to manage.^[1]

How an agent system sequences its steps, tools and models: a fixed chain, routing to specialized handlers, parallel work, or an orchestrator model that hands subtasks to worker models and combines their results.^[2][1]

A limited trial of an agent, often on selected cases. A pilot shows feasibility; only results on live volume over time show a production return.^[3]

The point at which an agent can run on live volume without rescue: tested against real cases with known answers, permissions limited to what it needs, oversight and logging in place, monitoring running, and an owner named to maintain it.^[10][6]

An attack where instructions hidden in content the agent reads, such as a document, web page or email, try to change what it does. OWASP ranks it first among risks to LLM applications.^[6]

Running an agent alongside the existing process without using its output, so people can compare its work with their own before trusting it.

The functions an agent can call to read data or take actions, such as searching a database, updating a record or sending a message. OpenAI's guide lists tools with the model and instructions as an agent's three core parts.^[1]

A repeatable sequence of steps that turns a trigger into a finished output, such as an inbound document into a filed record. In Anthropic's usage, an LLM workflow follows predefined code paths, unlike an agent that chooses its own steps.^[2]