I would say the first starting point is to run your agent somewhere you're comfortable with giving it mostly unconstrained permissions (e.g. --dangerously-skip-permissions for Claude COde), but more importantly, setting up sub-agents to hand off most work to.

A key factor to me in whether you're "doing it right" is whether you're sitting there watching the agent work because you need to intervene all the time, or whether you go do other stuff and review the code when the agent think it's done.

To achieve that, you need a setup with skills and sub-agents to 1) let the model work semi-autonomously from planning stage until commit, 2) get as much out of the main context as possible.

E.g. at one client, the Claude Code plugin I've written for them will pull an issue from Jira, ask for clarification if needed, then augment the ticket with implementation details, write a detailed TODO list. Once that's done, the TODO items will be fed to a separate implementation agent to do the work, one by one - this keeps the top level agent free to orchestrate, with little entering its context, and so can keep the agent working for hours without stopping.

Once it's ready to commit, it invokes a code-review agent. Once the code-review agent is satisifed (possibly involving re-work), it goes through a commit checklist, and offers to push.

None of these agents are rocket-science. They're short and simple, because the point isn't for them to have lots of separate context, but mostly to tell them the task and move the step out of the main agents context.

I've worked on a lot more advanced setups too, but to me, table stakes beyond minimising permissions is is to have key workflows laid out in a skill + delegate each step to a separate sub-agent.