> agentinfra_/glossary
Plain-language definitions of the AI agent and infrastructure terms used across the leaderboard.
AI agent
An AI agent is a system that uses a large language model to plan, reason, and act toward a goal on its own: deciding what to do, calling tools, and completing tasks with limited human oversight.
agent framework
An agent framework is a developer library for building AI agents — providing primitives for tools, memory, loops, and error handling so teams don't build agent plumbing from scratch.
LLM inference
LLM inference is the process of running a trained large language model to generate output — the compute-heavy step that model providers serve as APIs and that inference servers optimize for speed and cost.
RAG
RAG (retrieval-augmented generation) gives an LLM access to external knowledge: a query retrieves relevant documents from a vector or search index and feeds them to the model as context, grounding answers in private data and reducing hallucinations.
MCP
MCP (Model Context Protocol) is an open standard for connecting AI agents to tools and data sources — one common protocol so any agent can talk to any server exposing capabilities like file access, APIs, or databases.
agent orchestration
Agent orchestration coordinates multiple agents, steps, and LLM calls into one workflow — managing state, retries, and handoffs so a complex task runs reliably end to end.
evals
Evals are automated tests that measure how well an AI agent or LLM performs — scoring outputs against expected results to catch regressions before they ship, covering quality, safety, and cost.
agent memory
Agent memory lets an agent retain information across turns and sessions — short-term context, long-term stores, and recall that let agents personalize and continue long-running tasks.
prompt injection
Prompt injection is an attack where hidden instructions inside tool output or web content override the agent's original instructions — the top security risk for AI agents, defended by guardrails and output filtering.
computer-use agent
A computer-use agent operates software the way a person does — moving a cursor, clicking, and reading screens — automating web and desktop tasks without APIs.
voice agent
A voice agent conducts spoken conversations using speech-to-text, an LLM for response generation, and text-to-speech — powering phone lines, apps, and devices with real-time dialogue.
hallucination
Hallucination is when an LLM confidently produces false or fabricated information — mitigated with RAG, citations, and evals, and especially risky in agent outputs because they flow into actions.
tool use
Tool use (function calling) lets an LLM request external actions — API calls, searches, code execution — and incorporate the results, which is what turns a chatbot into an agent.
agent observability
Agent observability tracks what an agent did and why — logging every LLM call, tool use, and decision to debug failures, audit behavior, and measure cost per task.