Chatbots answer. Agents act: browse, run code, send mail, edit files.
That changes the UX problem. You need to show what the agent is doing, let people interrupt, and decide when a human takes over.
Six patterns keep showing up in products that do this without scaring people off.
1. Plan-execute
Before the agent runs, show the planned steps. People see what will happen, not only what already did, and can stop expensive or irreversible work early.
Interactive Demo: Plan-Execute
I broke your goal into 4 steps. Review the plan, then run when ready.
Plan ready
Friday customer demo
Booking a flight, sending mail, deploying code: hard to undo. A plan first turns a black box into a proposal you can reject.
→ Explore the Suggest / Confirm / Execute pattern
2. Visible tool use
Agents call APIs, query databases, run code, browse the web. Show which tool ran, what went in, and what came back.
Interactive Demo: Tool Use
When something breaks, people can tell if the failure was reasoning or a bad tool call.
→ Explore the Tool Use pattern
3. Task queue
Big goals split into subtasks that may run in parallel. A queue with pending, running, completed, and failed answers the only question people always have: what is it doing right now?
Interactive Demo: Task Queue
→ Explore the Tool Use pattern
4. Workflow builder
Not every run should be improvised. Sometimes people want fixed triggers, conditions, and steps the agent must follow. That sits between full autonomy and old-school automation, and it is what compliance teams usually ask for.
Interactive Demo: Workflow Builder
→ Explore the Workflow Builder pattern
5. Human handoff
Define when the agent stops and a person takes over: low confidence, high stakes, or an explicit ask. The failure mode to avoid is confident action in an uncertain situation.
Interactive Demo: Human Handoff
I can see two charges of $49 on Mar 3. I can open a refund request, but billing disputes need a person to confirm.
Connect you to billing support with this chat history?
→ Explore the Human Handoff pattern
6. Error recovery
APIs time out. Data is missing. Assumptions fail. Good recovery UX suggests a retry with changes, or another path, instead of a dead-end error.
Interactive Demo: Error Recovery
Ready
Pull current inventory for SKU-1842
On failure: retry up to 3 times, then escalate.
EVENT LOG
No incidents yet.
→ Explore the Error Recovery pattern
Calibrating trust
Too little trust and the agent sits unused. Too much and people get burned. Aim for confidence that matches how reliable the agent actually is.
The agents people delegate to are rarely the most autonomous. They are the ones people can see, stop, and hand off.