A set of interaction patterns for delegation, review, interruption, and human override.
- Year
- 2026
- Focus
- Patterns for delegated work
- Role
- Interaction design, agent workflow modelling, prototyping, pattern definition
- Themes
- Agentic workflows, Human control, Interaction patterns, AI product design
How can a person stay in control when work is being delegated to an agent?
Case specimen
Patterns for delegated work
- Messy inputs
- Delegation scope · Agent progress · Review checkpoints · Override moments
- Working frame
- Define handoffs as explicit states: assign, clarify, monitor, review, interrupt, override, and resume.
- Decision enabled
- Help users decide when to let the agent continue, when to intervene, and how to preserve context across handoffs.
- What changed
- The kit created reusable patterns for making agentic workflows feel inspectable and interruptible.
The situation
Context
As AI products move from answering questions to carrying out tasks, interaction design has to support delegation, progress visibility, correction, and accountability.
Problem
Many agentic interfaces blur the line between what the user asked for, what the agent is doing, and when the human can step back in.
The work
My role
I modelled agent states, designed handoff moments, and prototyped patterns for review, interruption, override, and resumable context.
What made it hard
The hardest part was giving people control without forcing them to micromanage every step of the agent's work.
Process
I mapped the lifecycle of delegated work, identified failure and uncertainty moments, and translated them into reusable interaction patterns.
Key design decisions
The kit used explicit task scope, progress checkpoints, visible assumptions, interruption controls, and review states that keep the human decision role clear.
What it changed
Outcome
The work produced a practical pattern set for designing agentic workflows where autonomy remains bounded by human judgement.
What I learned
Agentic products need control surfaces as much as they need conversational surfaces. The handoff is the interface.
What I would do differently
I would run scenario testing with users across low-risk and high-risk tasks to tune when interruptions should be prominent or quiet.