NeuralAgent Now Uses Your Computer At Lightning Speed

@kate_sleeman Great question, this is core to making it usable. NeuralAgent doesn't just fire actions blindly: a reasoning model supervises execution, so when an action doesn't produce the expected result, it can detect that and recover or replan rather than charging ahead. The fast model handles the rapid execution; the supervisor is there to catch and correct. It's an area we're continually hardening, mistake recovery is honestly one of the hardest and most important parts of computer-use, so we treat it as a first-class problem, not an afterthought.

@ill_robyn Thank you! Reliability on complex workflows is exactly where we focus most. Two things help: NeuralAgent plans the task and a reasoning model supervises while the fast model executes, so if a step goes wrong, it can catch it and recover/replan rather than blindly continuing. And for workflows you repeat, you can save a successful run as a Replay so it runs the proven sequence reliably each time. Still actively improving on the hardest multi-step cases, would genuinely love to hear what workflows you'd want to throw at it.

@christian_onochie You're thinking exactly right and we built for this. NeuralAgent has two permission modes: Ask First, where it pauses and asks for your confirmation before sensitive actions, and Act Without Asking for when you want it to run uninterrupted. So you stay in control of what needs a human check versus what runs freely. Appreciate you raising it, control over sensitive actions is non-negotiable for this kind of product.