Human-Approved AI Plans

Coding agents are most useful when they can act, and most dangerous for the same reason. Human-Approved AI Plans is the VibeControls concept for putting a readable, reviewable artifact between an agent's intent and its effect on real systems. Rather than watching a transcript scroll past and hoping to interrupt in time, a reviewer sees the whole sequence first, comments on the steps that worry them, and holds the only control that starts execution.

The first illustration places plan mode inside the AI workbench. The plan is a numbered checklist for a data migration: snapshot a table, add a nullable column, backfill rows in batches, move a read path behind a feature flag, then drop the legacy column after a bake period. Each step has its own checkbox so a reviewer can work through the plan line by line; in the sample the first three are ticked and the last two are still open. A side rail gathers annotations from named reviewers — in the sample data one asks to hold the backfill until off-peak hours, another asks for a rollback note on a specific step — and a single prominent approve control sits at the foot of the panel. A header strip names the agent the plan would run on and reports the connection and the hooks configured for it.

The second illustration extends the same review pattern from plans a person requested to proposals the platform raises on its own. A queue of cards describes candidate work — recovering a degraded agent, reclaiming idle sandboxes past their lifetime, reconciling drifted catalog entries, triaging unowned findings — each tagged by category and marked as awaiting approval. Selecting a card opens its full step list, a reviewer note, an estimated credit cost and the model that would do the work, alongside approve and deny controls and a plain statement that nothing runs until the plan is approved. A banner across the design marks this proposal engine as a forward-looking idea rather than a settled surface.

Together the two screens describe a governance shape the rest of VibeControls reuses: plans are objects, not chat. They can be read by someone who did not write the request, annotated by more than one person, costed before they run and tied to the audit trail afterward — which is what lets a team give agents more reach without giving up the review step.