Delegation to artificial intelligence in intellectual work should maximize human epistemic agency rather than minimizing effort

As more intellectual and technical work is delegated to autonomous AI agents - sometimes many at once - it is tempting, especially in non-critical applications, to delegate as much as possible and simply iterate: let the agent review and refine its own output until the result looks satisfactory. However, this does not necessarily converge reliably. Some errors persist however long the agent run is, whether carried out by a single agent or several working together on their own output, and current practice has no principled answer to the two questions this raises: where does AI autonomy actually reach its limit, and how should human review be organized to stay minimal without missing what matters? I argue that letting assignment specification be discovered through interaction instead of having it supplied upfront partly answers the second question, with the optimal loop being the one that maximizes human epistemic agency instead of simply minimizing human effort. Toward the first, I propose a decision rule for how tightly a human - AI loop should be coupled as a function of a task's novelty and the cost of reversing a wrong step. I support the position with three case studies - AI-assisted academic writing, mathematics, and agentic assignment execution - address alternative views on upfront specification and capability scaling, and close with implications for authorship and for how AI use may reshape human epistemic capability over time.