Most teams design agent systems as if the user only meets the final answer.
That is usually wrong.
The user also meets the wait, the retry, the "still working" state, the escalate-to-human moment, the half-done draft that needs a choice, and the moment the system says it cannot continue.
Those are handoffs. In a multi-agent product, the handoff is the interface.
If you only polish the final file, you are decorating the exit while the hallway is dark.
What users actually feel
Watch someone use an agent flow for real work: research, brief, draft, review.
They do not feel "eight clever agents." They feel:
- how long a step can stall before trust dies
- whether the system names what it is waiting on
- whether incomplete work arrives as a clear ask or as confident mush
- whether escalation feels designed or broken
- whether they can resume without starting over
That list is product design. If the hop is vague, the screen will be vague. Same seam problem as agent chat without contracts, now on the surface the user sees.
Name roles later. Design the seam first.
Teams love naming agents: Researcher, Planner, Writer, Critic. Naming is easy.
The work is the seam between Writer and Critic.
For each handoff the user can see, write:
- State - what is happening, in plain words.
- Time - how long it should take, or an honest "unknown."
- Action - wait, approve, edit, choose, or cancel.
- Draft - what they can inspect before the hop finishes.
- Escape - how a human can take over without restarting from zero.
No escape means you did not design autonomy. You designed a trap.
That is why escalation comes before more agents. Escalation is not an apology screen. It is part of the product language.
Confident mush is a UX bug
Agents sound finished even when they are unsure.
On a product screen, that is dangerous. A guess dressed as a final answer trains people to stop checking. Then the rare expensive failure feels like betrayal.
Better pattern:
- mark verified versus proposed
- show what passed or failed the handoff check
- keep open questions visible as open questions
- never let "final" mean "nobody looked"
Make honesty cheaper than vigilance. If only engineers can see the truth in a log, the product is lying politely.
Memory shows up in the handoff too
People hate repeating themselves. Teams dump the full history into every agent.
Then agents inherit old assumptions and private scratch, and speak as one clean voice. The user sees coherence. The system has washed confusion.
Pass only what the next role needs. That is workflow memory done on purpose, felt as continuity instead of amnesia.
What this is not
Not "add a spinner and call it agent UX."
Not "show the whole chain of thought to prove intelligence." People need decision-grade state, not a novel.
Not "every hop must be visible." Quiet hops are fine when failure is cheap. Visible design belongs where trust, money, brand, or irreversible action sits.
Do this on Monday
Map one live flow as the seams a user can see. Fill the five fields for each seam. Run it once.
Note where the UI lies: silent waits, false finals, missing escapes, or resumes that force a full restart.
Fix the worst seam before you add another agent. New agents multiply seams. Seams multiply product surface.
Take-home
- Users meet agent systems at handoffs, not at model cards.
- Design state, time, action, draft, and escape for every important seam.
- Treat confident mush as a product bug.
- Pass only the memory the next step needs.
Related reading:
- Agent Conversations Without Contracts Are Faster Confusion
- Design the Escalation Layer Before You Add More Agents
- Designing Workflow Memory Before You Add More Agents
- Agents Should Not Share a Brain
If you want agent seams designed as product, not bolted on after the demo, book a discovery call.