An agent interface needs an abort button

An agent interface should expose state without forcing the operator to tail logs all day: what is running, what changed, what is blocked, what requires approval, and what can be interrupted without corrupting the work.

The runtime already knows most of this state, but interfaces flatten it into “thinking” or a stream of tokens. Lifecycle states need timestamps and ownership: queued, gathering context, executing a tool, waiting on a human gate, verifying, completed, failed, or cancelling. Background work needs durable completion notices reconciled against live state so a delayed notification cannot trigger a stale retry.

The control surface is not chat-only. It needs threads, run state, approvals, memory, diffs, stop controls, and a fast path back to the current objective.

Each approval should name the exact command or side effect, scope, affected resources, reversibility, and the state that will follow acceptance or rejection. The diff and verification results belong beside the approval rather than in a separate log hunt. A stop control must say whether it requests cooperative cancellation, kills an owned process, or leaves an external operation potentially ambiguous.

Autonomy earns trust through inspectability. The moment a system cannot explain its state or accept an abort, the operator has lost the advantage.

The interface should optimize for exception handling, not celebratory demos. Healthy runs collapse into concise evidence. Stalled loops expose repeated actions and budget. Partial failure preserves artifacts and a resumable packet. Human input gates surface one concrete decision. This lets the operator supervise several durable workflows without becoming the scheduler, while keeping authority over destructive or externally visible actions.