# Act, Confirm, or Stop? Smarter behavior for AI assistants, wearables & robots — Amit Desai, Roku

Source: [AI Engineer](https://www.youtube.com/watch?v=Zd5b40Jbp_k)  
Feed7 permalink: https://feed7.dev/p/act-confirm-or-stop-smarter-behavior-for-ai-assistants-wearables-robots-160uz4o  
Published: 2026-09-15T17:00:32.000Z  
Trust: Source Linked (source_linked)

## Why Included

An assistant can reduce user pain without improving model accuracy by choosing when to act, confirm, or stop. Optimize those decisions against the cost of each failure mode, not intuition.

## Source Summary

The example holds recognition accuracy at **79%** and assigns user-cost estimates to wrong actions, stops, and confirmations. Immediate action costs 2.1 points per turn; a guessed 65% stop threshold lowers that to 1.904 but is not optimal.

## Practical Implication

Model act, confirm, and stop as separate decisions under uncertainty. With the stated costs, optimization selects **41% and 49% thresholds** and lowers the modeled cost to **1.464 points per turn**, without changing the underlying predictions.

## Agent-Ready Context

The example holds recognition accuracy at **79%** and assigns user-cost estimates to wrong actions, stops, and confirmations. Immediate action costs 2.1 points per turn; a guessed 65% stop threshold lowers that to 1.904 but is not optimal.

Model act, confirm, and stop as separate decisions under uncertainty. With the stated costs, optimization selects **41% and 49% thresholds** and lowers the modeled cost to **1.464 points per turn**, without changing the underlying predictions.

Those thresholds are not portable defaults. The cost values are heuristics, and they change with modality, recovery time, and consequence; sending an email or moving a robot warrants different penalties than playing a song.

## Connected Context

Feed7 judgment across 812 accumulated Signals:

This isolates runtime decision policy as a reliability lever: unchanged predictions can cause less user harm when acting, confirming, and stopping are optimized separately. It makes safe interruption quantitative rather than intuitive, while narrowing the result to chosen cost assumptions that must be recalibrated for each interface and consequence profile.

- [From RL to IRL — Gaurav Mishra, Amazon AGI Lab](https://feed7.dev/p/from-rl-to-irl-gaurav-mishra-amazon-agi-lab-0ex196x) — The risk-aware checkpoints and user handoffs described for UI failures provide concrete places to apply cost-calibrated confirm-or-stop decisions.
- [Selective Agent Guidance via Entropy: Learning Autonomous Policies from Imperfect VLM Teachers](https://feed7.dev/p/2609-01567v1-0mj0nht) — Both gate behavior on uncertainty, but SAGE uses uncertainty to request training-time guidance whereas this Signal uses it to choose among runtime action, confirmation, and stopping.
- [Desktop-Delta Bench: Do Computer-Use Models Understand Desktop GUI Transitions?](https://feed7.dev/p/2607-26041v1-1x1gw81) — Step-level evidence about GUI transitions could inform whether continued action is safe, complementing the policy for deciding when to act or interrupt.

## Context Map

- Layer: agent
- Domains: audio
- Topics: agent-reliability, computer-use

## Uncertainty

- Those thresholds are not portable defaults. The cost values are heuristics, and they change with modality, recovery time, and consequence; sending an email or moving a robot warrants different penalties than playing a song.

## Agent Instruction

Use this item as source-backed context. Do not invent claims beyond the linked source. If this item conflicts with another source, call out the conflict.