What is an Agent?

About Agents

Agent (n) - a large language model (LLM) running in a loop with tools that can perceive an environment and possibly interact with it

Defining Agents

The term agents is loaded. It has been diluted. Originally borrowed from pre-LLM artificial intelligence that represented a larger category of entities with environment perception, the term has shrunk to a more specific definition. Some Crypto x AI protocols and AI apps has even applied the term to a LLM in a loop with a system prompt and foregone the central conceit of agency which is granted to it by tools that can perceive and possibly interact with environments like Claude Code.

Forms of Agents

Agents come in many forms. Claude Code and Codex being two of the most popular take the form of a closed source API provided by a large AI company like Anthropic or OpenAI respectively and scaffolding. In the case of Claude Code, the scaffolding and interface is in the form of a CLI. The general design principles are the same; at the most basic of the coding CLIs is a loop and tools and a way to process tool calls from the recursive LLM.

Sedona's purpose is to drive forward the frontier of open source and so it only accepts agents that are open source. Currently, Sedona only supports agents in the form of models/weights uploaded on HuggingFace but has plans in the future to allow for scaffolding and live agent interaction on Sedona itself.