Does user editing of AI plans always improve performance?

The same empirical study (N = 248) of plan-then-execute LLM agents found that letting users intervene helps task performance overall, but the benefit is stage-dependent and, at the planning stage, quality-dependent. On planning: "user involvement in planning can help address imperfect plans (e.g., task-1 with grammar error)," which "further contributes to improvements in the execution accuracy." But the same intervention cuts the other way when the plan starts out good: "user involvement in planning can harm plan quality in tasks with a high-quality initial plan, which may potentially cause worse task performance in the subsequent execution stage." On execution: after controlling for plan quality, "user involvement in the execution can provide the best execution accuracy among most tasks considered in our study" — a more stable, direction-consistent benefit than planning-stage editing produced.

The source attributes the asymmetry to what each stage's interface lets users do. Planning-stage editing is a global rewrite of the step-wise plan via generic edit tools, with no per-step scaffolding to tell a user whether a given plan segment is already correct — so an edit can introduce as much damage as it fixes. Execution-stage review is local and bounded: for each primary step, users see one predicted action and choose among three options — "Proceed" (accept), "Feedback" (request a new prediction), or "Specify Action" (manually pick from a fixed, pre-defined action list and fill parameters) — checking a single concrete action against known failure types the authors name as "prediction errors (i.e., wrong action name or action parameters) and prediction failures (i.e., failure to provide valid action prediction)." A bounded check against a known failure type is lower-risk than an open-ended rewrite of free-form plan text.

This is the same general move as Why do planning and grounding pull against each other in agents?, but discovered on the human side of the interaction rather than inside model training: AutoGLM separates planning and grounding because the two need opposite optimization pressure during training, while this study finds that user oversight of planning and oversight of execution need opposite interface designs because free-text editing and bounded action-selection carry different risk profiles for a human reviewer. It also sharpens what "human control" should mean in What UX principles do workplace users want in AI agents?: control exercised as plan-editing is not uniformly protective, while control exercised as execution-step review is. And it adds a performance side to Should AI systems stay collaborative rather than fully autonomous? — human involvement does help, but the excerpt shows that help is not automatic at every stage of a workflow.

The excerpt does not say how often plans started out high-quality versus low-quality, nor whether users could tell in advance which kind of plan they were editing — the study does not test a design that signals plan confidence before inviting an edit. The implication the source draws, and the excerpt supports at this strength, is that a plan-review affordance is not free: offering it unconditionally risks degrading plans that didn't need a human hand, and interfaces that invite open-ended plan edits should either gate that invitation on some quality signal or default review effort toward the bounded, step-by-step execution check instead.