Make plans visible
Users trust agents more when they can see the proposed steps before irreversible actions. Show the plan, allow edits, then execute.
Separate tools from judgment
Give agents constrained tools with clear side effects. Keep high-risk actions behind confirmation. Logs should show which tool ran, with which arguments, and why.
Design for interruption
Long-running agent work must be pauseable and resumable. People leave mid-flow. State should survive that reality.
