Autonomy and Control With Coding Agents

One of the harder parts of working with coding agents is not technical. It is deciding how much of the coding process I still want to be inside.

Engineers are landing in very different places. Some want to review every line. Some behave more like product managers: define the outcome, check that it works, and let the implementation happen elsewhere. Some want an employee. Some want a coworker. Some mainly want the job done.

I do not have a lesson here. I am still looking for the workflow that gives me the right balance between autonomy and control. Every now and then I think I have found it, then I learn about another tool, model, or workflow and the research starts again.

That balance is personal. Running three or four parallel worktrees can feel fast, until the end of the day when you realize you mostly compared outputs. Delegating deeply can feel productive, until two weeks later someone asks about a detail and you notice you did not build the same internal map you normally would. Maybe the answer is to ask the agent. Maybe the answer is to stay closer next time.

It reminds me more of choosing an IDE, a terminal setup, or Emacs over Visual Studio than picking a universally correct methodology.

Try tools, models, and workflows yourself. Timebox the experiments. Notice what you gain, what you lose, and what kind of work still feels like yours.