Glossary · Machine learning
Coding agent
Also called AI agent, agent, AI harness, agent harness
An AI assistant that doesn't just answer questions but writes code, runs it, reads the results and keeps going, on a real computer.
A chat assistant answers in text. A coding agent is a large language model connected to tools: it can read and write files, run programs and look at what they print. Given a goal, it works in a loop: plan a step, take it, check the result, decide the next step. The person sets the direction and checks the work. The agent does the typing.
The software around the model that gives it those tools, keeps its session going and lets a person talk to it is often called the harness. One harness can be reached from more than one place, for example a terminal and a phone app connected to the same session.
What it’s good and bad at
An agent is fast at the mechanical work: writing a reward term, launching a run, reading thousands of log lines, making a video. It is much worse at judging whether a simulated jab looks like a jab. That judgement still needs a person, which is why the two work as a pair (see human-in-the-loop).
In this project
The coach’s agent runs on a small always-on server and he talks to it from his phone. It builds reference motions and sends back pictures and short clips, writes the environments and rewards, launches each training run on a cloud machine, watches its logs, sends grids of test videos to the phone and suggests what to change next. The coach decides.