Can users steer AI agents better by revising plans mid-execution?

Cocoa, a document-editor agent system for scientific researchers, lets users assign each step of an agent-proposed plan "to either themselves or the agent" and then interleaves planning and execution so "plans can adjust after partial execution" rather than treating the two as fixed, separate stages. The paper argues most prior systems either elicit user intent only before execution or let users verify output only after it, while Cocoa supports "co-planning," "co-execution," and smooth movement between them, since "plans are bound to shift as the user reviews agent outputs and becomes inspired with new ideas." A lab study (n=16) found Cocoa "afforded greater agent steerability while maintaining ease of use" against a chat baseline, and a 7-day field deployment (n=7) found researchers valued "the explicit delegation of work between themselves and the agent."

The design borrows "inspiration from computational notebooks": plan steps live as cells in the document that either party can execute, re-execute, or hand off, so delegation is visible and reversible rather than a one-time setup choice. The paper frames this as resisting "agent-guided workflows" where "the agent dictates what input to elicit from the user, as well as when and how the elicitation happens" — Cocoa instead lets the user set per-step agency. The discussion names design tensions the system does not fully resolve: re-planning currently "replaces outdated steps with new ones," though participants wanted to preserve superseded steps for later reference, which the authors propose addressing through "plan branching"; and plan length should be calibrated to task type, since open-ended brainstorming wants short, revisable plans while a structured task like literature review wants a full plan upfront.

This extends the stage-dependent finding in Does user editing of AI plans always improve performance?, which treats planning and execution as two fixed interface moments with different risk profiles for human review; Cocoa's contribution is a design that removes the fixed boundary itself, letting users move between planning-edits and execution-edits as often as a task demands rather than choosing once which stage to intervene in. The paper cites He et al.'s finding that separating planning from execution lets user involvement during execution "significantly boost agent performance" but at "high cognitive load" — a tension also visible in Can orchestration layers make coding agents more auditable?, another design that keeps the human at a defined checkpoint rather than interleaving — while arguing the fix is not to avoid interleaving but to make hand-offs between planning and execution explicit and assignable per step.

The excerpt does not report how often users actually assigned steps to themselves versus the agent, nor whether the steerability gain holds for tasks less structured than scientific research, since both studies drew participants only from that domain. The implication the authors draw, and the excerpt supports only at this scope, is that interaction designs for long-running agent tasks should make delegation a per-step, revisable choice rather than a global setting — not that any particular assignment ratio or plan length generalizes across domains.