Models, Agents, Assistants, and Harnesses
This article explains some of the jargon around AI and agents.
If you are new to AI, you may come around a lot of jargon and not know what they all mean. One of the common confusions when getting started is related to models, agents, assistants and harnesses.
Model (brain)#
Model is the brain. It is the raw neural network (GPT-4, Opus, Llama, Gemma) trained on data to predict and generate text/tokens. Their function is: text in, text out.
A model has no memory, no access to tools, and no ability to act on the world. It only produces a response based on its weights and the input it is given.
Harness (scaffolding)#
A harness is the software scaffolding that turns a model into an agent or assistant. It provides prompt construction, tools, context, execution, memory and permissions. The model "thinks"; the harness wires it to the real world, feeds results back, and enforces the rules. Claude Code is a harness, so is OpenClaw.
You can take different models (Gemma, Llama, Opus, GPT-5..), and plug them into different harnesses (Pi, OpenCode, Hermes, OpenClaw..). Based on what harness you use, you can achieve different things with the same model.
Some harnesses are open and free and allow you to use any model (e.g. OpenCode, Pi, OpenClaw, Hermes) while others force you to use only a limited set of models (e.g. Claude Code can only use models by Anthropic)
Agents and Assistants#
Agents and assistants are both harnesses. The differentiation between agents and assistants is not that strong anymore as both exist on a spectrum of increasing capability and autonomy. You might even see the terms used interchangeably. OpenClaw calls itself an assistant while Hermes calls itself an agent.
Assistants are a user-facing product optimized for chatting with a human. They are generally reactive, meaning it waits for you to ask before it does anything.
Agents are given a goal, autonomy, and the ability to act in a loop. It plans, calls tools (search, code, APIs), observes the results, and iterates until the task is done, without a human prompting each step. The key difference from an assistant is autonomy and tool-driven action towards a goal, not just answering.
Claude Code, Codex, OpenCode and Pi are all coding agents.