AI definitions
Definitions around AI can vary and hallucinate based on their context, so I’m assembling what I found in the wild to make sense of it for myself.
This is a quick reference for the terms used to describe AI agents. Definitions are gathered from various sources.
Agent
A large language model (or models) (LLM) running inside a harness, able to complete tasks on its own rather than only producing text.
Harness
The harness is what the LLM uses to navigate the world and get work done. It gives the agent memory, tools, and a safe place to act. The environment around the LLM that supplies memory, tools, and sandboxing so the agent can act with real autonomy.
What’s inside a harness:
- Filesystem — lets an agent store and retrieve its work so progress persists.
- Tools — let an agent take action instead of only describing what should happen.
- Sandbox — the bounded environment where an agent’s actions, such as file operations and shell commands, run safely.
- Context — the information and configuration given to an agent so it stays effective across long tasks.
- Subagents — let an agent split a job and work on several tasks at once.
- Headless agents — agents without a built-in user interface. They run through APIs, not a chat window or CLI.
- Programmable — agents are built and run entirely in code: you set the model, instructions, tools, and environment.
- Session — an instance of an initialized agent that you send prompts to and receive responses from.
- Skills — bundles of expertise an agent loads on demand to extend what it can do.
Sources