The LLM is the pilot, not the plane
A better model makes an agent smarter. But producing a better answer is not the same as achieving an outcome.
What turns an LLM into an agent
An LLM can reason and generate. To actually complete a task, it needs a system around it that gives it the right information, lets it take actions, and shows it what happened so it can decide what to do next. This system is often called the agentic harness.
Think of the LLM as the pilot inside an agentic system. The agentic harness is the cockpit around it: it provides the instruments to understand what is happening, the controls to take action, and the feedback needed to decide what to do next. The pilot provides the intelligence and makes decisions, but it is the complete system that allows us to reach the destination.
In practice, a typical agentic loop looks like this: reason → act → observe → repeat. This is how an agent moves from producing an answer to pursuing an outcome.
Coding agents are not only for software development
Coding agents make the importance of the harness especially clear. They operate inside a programmable environment where they can inspect files, run commands, use existing tools, write programs, and observe the results. This gives them the ability to adapt how they approach a task rather than relying only on a fixed set of predefined actions.
Consider an invoice that needs to be converted into a specific format. A coding agent can inspect the document, choose an existing tool or write the code it needs, transform the file, check the result, and adapt when the input or the output has an unexpected structure.
A coding agent does not have to produce software. Programming is simply one of the means it can use to achieve an outcome.