Apr 2026
Designing agents people can steer
Design Notes Design
Stepping off the loop only works if you can still see it and steer it. Visible progress and reviewable evidence are what make that safe.
An agent that disappears for a minute and returns with a result is hard to trust and harder to correct. If people are going to step off the loop, steerability is what makes it safe — and that is a design problem, not just a model property.
We design Offloop agents so a person can step away, glance back at any time, and step in without starting over.
Show the work as it happens
Visible tool activity lets a person glance at what an agent is doing while it is still doing it, instead of judging only the final output.
Make stepping back in cheap
Re-joining a loop should not mean throwing away progress. Corrections land inside the same task so the agent adjusts rather than restarts.
"Stepping off the loop is only comfortable when you can step back onto it at any moment."
Leave evidence, not just answers
Reviewable artifacts let a person accept or reject a result on its merits — which is what keeps control in human hands even when nobody is watching the loop.
Design Notes writes product notes for Offloop, focusing on agent workflows, persistent Channels, and the craft of keeping delegated work inspectable.
