Definition:
An A2UI is an emerging interface paradigm designed for collaboration between autonomous AI agents and human users. Unlike traditional UIs where a person must know exactly where to click and what to type, an A2UI treats the agent as an active partner: it proactively shares its plans, exposes its reasoning, requests permission at critical check-points, and reports progress in real time. The conversation flows both ways—Agent-to-User, not just User-to-System.
In short, if a GUI is a dashboard, an A2UI is a copilot’s cockpit.
From Commanding to Collaborating Traditional UI / API A2UI User knows the goal and every step. Agent proposes the goal and a plan. Error messages appear after failure. Agent flags risks before acting. Static forms and buttons. Dynamic cards showing reasoning, options, and progress streams. “Click to execute.” “Here is what I’m about to do. Approve, modify, or stop me.”
Anatomy of an A2UI Element Purpose Plan Transparency Shows the agent’s intended steps so the user can edit or remove them before execution. Context Gate Asks the user to fill a missing gap the agent cannot infer with certainty. Approval Checkpoint Pauses the workflow when stakes or costs are high, awaiting human consent. Progress Streaming Displays live status as the agent works across tools or long time horizons. Undo / Fork Lets the user roll back or branch a plan if the agent’s output diverges from intent.
Example:
“Clara,” the Scheduling Agent Imagine your company deploys Clara, an agent that manages your calendar.
The Traditional Way
You open your calendar app, compare your schedule with three coworkers, find a free conference room, paste in the Zoom link, and send invites. Clara is just a dumb form filling in the background while you do the mental work.
The A2UI Way
You casually message Clara: "Set up the quarterly review with the Design team this week."
Clara responds through her A2UI panel:
Plan Transparency
Clara shows a card:
“I’ve scanned 5 calendars. I find a conflict: the only shared free slot is Friday 2 PM, but you are booked to fly out that afternoon. Here are three paths I can take:”
A. Book a 30-min virtual meeting for Monday 9 AM.
B. Ask Design to move their sprint review so we can do Thursday 1 PM in Room 302.
C. Schedule for Friday but shorten it to 15 mins with a mobile link.
You tap Option B.
Context Gate
Clara asks:
“Room 302 requires catering for lunch meetings. Should I add a light lunch order, or is this a working session without food?”
You select “No food.”
Approval Checkpoint
Before sending, Clara displays:
“I am about to: move Design’s sprint review to Wed, book Room 302, generate an agenda from last quarter’s doc, and send invites to 8 people. Confirm?”
You hit Approve.
Progress Streaming
Once confirmed, Clara streams live updates in the sidebar:
✅ Room 302 reserved
✅ Sprint review moved (notified lead)
✅ Agenda draft attached
✅ Invites sent
“Done. I’ve also blocked 15 mins beforehand for you to review the Figma file.”
Follow-Up Context
Two hours later, Clara surfaces a new card:
“An attendee in another time zone noted a conflict. Shall I auto-record the session and send them the link?”
Result: You spoke one sentence. Clara handled the complexity but never acted blindly. Every significant decision was surfaced in an A2UI layer you could steer.