Assistant, agent, harness: what an AI agent really is

Assistant, Agent, Harness, Loop engineering, AI employee ... Now that everyone and their mother is busy with AI, the terms are flying past your head: agents, AI coworkers, AI assistants, Loop Engineering, Tool calling, harness... A lot of it is marketing buzz, so it is time for some clarity.

At BrA1n we notice in conversations with decision-makers that they are keen to get started with AI. But should we then have an AI assistant? Or an AI agent? What is an AI agent, actually?

An Assistant answers questions. Handy, but it does nothing by itself. So this is really an LLM, often with a thin shell around it that keeps the focus on a specific sector or organisation.

When an assistant is grounded, it gets interesting. Grounding means an AI assistant has access to one or more external sources. Think of the stock system, the calendar, a calculation module, or simply access to the internet. All of a sudden the assistant can actually look things up.

An AI agent is a different order of thing altogether. An AI agent as we define it has three main components:

1. An AI model, usually a Large Language Model (LLM). But it can also be several models for different tasks.
2. A set of tools. This is really an extended version of the grounding above. 'Access to the local computer' can be a tool too, or a Python interpreter to run code.
3. A harness. This is the "casing", or control layer, around the AI model. The harness composes the prompts that go to the AI model, analyses the answer, calls one or more tools, sends the results back to the LLM in a new prompt, and repeats that until there is something the user asked for.

That last part is what we call agentic AI: a system that gets to work on its own to complete a given task. This usually means many calls to the AI model, and therefore often a lot of tokens consumed. An agent typically runs for minutes to hours before the answer is ready.

In this context we increasingly come across the term loop engineering. This is the design of the working loop for an agent. A step higher up, in other words. The working loop determines when the agent may continue on its own and when a human needs to look at it.

Finally, AI employee or digital colleague: treat these as marketing terms. Such a term simply sells better than "language model with access to systems".