Agent IxD (Agent Interaction Design)
Coined by Sasha Merzliakov · First published August 2026
The design discipline covering how a person operates alongside one or more running AI agents. Direction is the interface for getting an instruction in, and verification is the interface for checking what comes back once the run finishes.
What does Agent IxD mean?
Agent IxD is short for agent interaction design. It borrows the IxD abbreviation the design profession already uses for interaction design, and narrows it to the case where the thing on the other side of the interface is an AI agent rather than a document, a form or a screen.
The distinction matters because an agent does not wait. It accepts an instruction, works for minutes or hours, and returns a result that has to be checked before it can be used. Direction and verification aren't sequential phases, they recur through a run. Briefing is direction. So is a mid-task correction. Verification shows up at checkpoints along the way, and again at the finished result. The split tracks what the human is doing, giving intent or checking output, not when in the run it happens.
Why Agent IxD matters
As models improved, the constraint moved. The expensive half of agent work is no longer the agent's own speed, it is the human either side of it: briefing it, redirecting it mid-task, and reading back what several of them produced at once.
Treating those as one undifferentiated "chat interface" problem hides the fact that the two halves have opposite properties:
- Direction is bursty, low-volume in characters, and latency-sensitive, so the interface question is how fast an intent can be got into the right agent
- Verification is high-volume, asynchronous, and judgement-heavy, so the interface question is which outputs need a human and in what order
- Speeding up one loads the other, because faster direction produces more agent output waiting to be verified
- The channels differ, and voice, keyboard, approval queue and diff view are not interchangeable across the two halves
Agent IxD and VUX
Agent IxD is the umbrella. VUX (Verification User Experience), also coined at hcd.ai, names the verification half of it: the design activities for checking what an agent team hands back. Voice, dictation and spoken routing sit on the direction half.
Human-Agent Interaction already covers much of this territory, alongside conversational UI, human-in-the-loop review and agent orchestration. Amershi et al. (2019), in Microsoft Research's CHI guidelines, split the same ground into 'efficient invocation' and 'efficient correction' seven years earlier. Google's PAIR guidebook (2019) covers the same pair under Feedback and Control, and Errors and Graceful Failures. Agent IxD doesn't claim new territory, and the direction/verification carve isn't new either. What it adds is a practitioner's working vocabulary for running an org of agents day to day, not a researcher's frame for evaluating one.
See the anchor article for the direction half in practice: Agent IxD: Voice.