i contributed to openai/codex, the open-source coding agent. codex reads a repository, uses terminal tools, works inside a sandbox, and leaves a diff a human can inspect.
the loop is simple enough to say out loud: the model proposes an action, policy decides whether it can run, the tool returns an observation, and the next step sees that result. the hard engineering lives around approvals, filesystem boundaries, logs, and the point where generated text becomes a real command.
agents get useful when they can act, and they stay useful when a person can reconstruct what happened after the run.