Plain definitions of common AI agent and automation terms.
A software system that pursues a goal across multiple steps, using a language model for reasoning and tools for actions, rather than producing a single reply.
The repeating cycle in which an agent observes the current state, decides on an action, runs it, and feeds the result back into the next step.
The degree to which a system can take steps toward a goal without a person approving each one; higher autonomy raises both usefulness and the need for guardrails.
The amount of text a language model can consider on a single turn; it is limited, so long tasks must carry forward only what matters.
A numeric representation of text that lets software compare meaning, commonly used to search a store of documents for relevant passages.
The practice of testing an agent against a set of known tasks to measure whether it completes them correctly and whether changes help or hurt.
Software that handles the repetitive parts of building an agent, such as the loop, tool calls, state and logging.
A rule or limit that constrains what an agent may do, such as a step cap, an allowed list of actions, or a required human approval.
Confident, fluent output from a model that is factually wrong; a key reason to attribute claims to sources and keep a person in the loop.
A model trained to predict text that supplies the reasoning and language ability behind most current agents; often shortened to LLM.
A design in which several agents with different roles pass work between each other; it can separate concerns but adds coordination cost.
The coordination of steps, tools, and sometimes multiple agents so that a larger task is carried out in the right order.
The instruction and context given to a model; clear, specific prompts generally produce more reliable agent behavior.
Fetching relevant information from an external store and adding it to the context, so an agent can use facts beyond its immediate window.
The mechanism by which a model requests an external action, such as a search or a calculation, and receives the result back.
A defined sequence of steps a task follows; automation runs a fixed workflow, while an agent can decide parts of the sequence itself.
FullOfAgents