Approval buttons show actions, not consequences: preview the trajectory before you commit

Gaole He, Brian Y. Lim · 2 authors · National University of Singapore · 2026 · arXiv:2603.11677

Approval buttons show actions, not consequences: preview the trajectory before you commit

Opening

Clicking approve tells the agent what may run now. It says nothing about where the run leaves the workspace a few steps later. That is the gap this paper starts from: interaction stays pointwise and reactive, and the long-term consequences of each approved action stay invisible to the person approving it.

What the paper does

This is a perspective paper. It characterizes the limitations of current interaction paradigms, introduces a conceptual framework for simulation-based collaboration, and illustrates its potential through concrete human-agent collaboration scenarios. Its proposal is simulation-in-the-loop: users and agents explore simulated future trajectories before committing to a decision, so intervention becomes informed exploration rather than reactive guesswork.

This perspective paper characterizes the limitations of current paradigms, introduces a conceptual framework for simulation-based collaboration, and illustrates its potential through concrete human-agent collaboration scenarios.

Findings

  1. Today’s approval loop is pointwise and reactive — users approve or correct individual actions to mitigate immediate risks, without visibility into subsequent consequences.

However, human-agent interaction remains pointwise and reactive: users approve or correct individual actions to mitigate immediate risks, without visibility into subsequent consequences.

  1. The missing foresight is charged to the user — pointwise approval forces users to mentally simulate long-term effects, a cognitively demanding and often inaccurate process.

This forces users to mentally simulate long-term effects, a cognitively demanding and often inaccurate process.

  1. Simulation-in-the-loop moves the decision before the commit — users and agents explore simulated future trajectories first, which also surfaces latent constraints and preferences along the way.

We propose simulation-in-the-loop, an interaction paradigm that enables users and agents to explore simulated future trajectories before committing to decisions.

Caveat

The paper offers a framework and illustrative scenarios, not evidence: no implementation, no user study, and no measured figures appear in the abstract. How accurate simulated trajectories are, and whether people can read them better than they guess today, is exactly what the paper does not establish.

What this means for 5ac

When we approve a card, the worker’s plan is a paragraph, and the consequences live somewhere after the merge. Change what we ask for before approval: the trajectory, not the promise. Have the worker show the branch diff, the files it will touch and the commands it will run, ideally produced by a dry run rather than described in prose. For cron and bot jobs, run the new job in shadow mode first, compare its output against the current one, then flip it on. Previewing the trajectory is cheap next to reverting a bad commit in a customer’s delivery.

Source