Prerequisites
Conversation has been chosen as this surface's primary mode, not one option among many. The agent's capabilities are listed, so there is a boundary to state: what it can act on, what it can only look up, and what it can only discuss. Both are settled before this sequence, and no pattern here makes them.
An actor and an AI agent meet in conversation, and the conversation carries the actor’s business. In a conversation-first interface the conversation is also the workspace, and structuring the space, composing views, and making changes are all reached from inside it.
Setting the footing
Draw the capability boundary — Capability & scope
State what the agent can act on, look up, or only discuss, and decide what it may promise. Give an out-of-scope request somewhere to go, never a dead end. Restate the boundary at each encounter, since capabilities change between visits.
Decide who can open
Most designs take both arms. Either way the actor has committed to nothing yet, so a way out before any activity starts is owed; the trouble steps build it.
Combine any of:
Let the agent open — Opening (Agent)
Open by saying who is speaking, what it can do, and what this is about, including the capability statement, and propose a first topic the actor can accept, replace, or turn down.
Recognise the actor's open — Opening (User)
Align to whatever the actor brought, and route it by shape: a question answerable as asked, a goal with steps behind it, or something too loose to act on, which the agent asks about first.
Carrying the business
Applies when
A topic is agreed, and the actor's business arrives in a few shapes: a question answerable as asked, a goal with steps behind it, or content too long for one turn.
Answer the askable directly — Inquiry (User)
Answer first, offer next moves after, and keep turns short. For anything non-trivial, attach the evidence to the answer.
Frame the goal-shaped request — Open request
Make the goal, constraints, assumptions, and sources explicit, so what is missing shows as named gaps.
Hold the floor for the long telling — Extended telling · optional
Run content that will not fit one turn as a structured sequence the actor controls: announce the scope, advance when they signal to go on, and keep skip, back, and stop available. The repair steps also hand over to this. Skip it where answers fit single turns.
Give turns a working surface — Inline interface · optional
Where words are the wrong vehicle, make the message something to act on: chips for a choice, a card to confirm, a chart to read. The confirmation and repair steps later use this surface. Skip it where the encounter stays in words.
Checking before the act
Applies when
Some requests end with the agent acting on its reading of them, by spending real work or making changes that are hard to undo, and the reading can be wrong. The check is decided per act, by its consequences, and routine acts go ahead.
Match the check to the act
Two checks, in no set order. Each act takes the one its consequences call for, and an act that can be cleanly undone takes neither.
In any order:
Preview the substantial work — Plan preview
Show the steps, targets, expected effects, and cost before substantial work, so a misreading surfaces while it is cheap. Preview only work that warrants it.
Take the go-ahead in place — Inline confirmation
For a single step that cannot be undone, ask for confirmation where the actor expressed the intent, without moving them elsewhere.
Designing for trouble
Applies when
Any activity can break in two directions: the agent misreads the request, or the actor does not follow the agent. Sometimes the actor simply wants out with the goal unmet.
Repair from both sides
Two directions of trouble, in no set order. Both resume the activity at the step that broke, not from the beginning.
In any order:
Show the agent's reading for correction — Agent repair
Show the agent's interpretation and parameters, and let the actor edit them. Where one specific value is missing, ask for that value instead of guessing again.
Re-deliver what did not land — User repair
Offer a paraphrase, a definition, or a worked example without the actor having to diagnose their own confusion. When none of them is enough, switch to a long telling.
Build the exit that is not failure — Abort
Let the actor stop an activity whose goal is no longer wanted: halt where it stands, state what changed, and keep what it collected visible for the actor to keep or discard. This is the way out owed since the opening. Changes already made are undone outside the conversation.
Ending and parking
Applies when
An activity has delivered, but only the actor's acknowledgement shows that it landed. The encounter as a whole has to be possible to end, and to pause without ending, with the difference visible to both sides.
Close the loop on each activity — Sequence completion
Treat the actor's acknowledgement, such as "thanks" or "got it", as the close of the activity. Offer a next move there only when the actor is likely to want one.
Settle the two endings
Design both endings, and let the kind of interface decide which one the defaults serve: a conversation-first interface ends deliberately and parks as the exception, and an embedded surface is mostly left open and rarely closed on purpose.
In any order:
End the encounter deliberately — Closing
Invite last topics and either handle or schedule them, so both sides know it is over rather than paused. Anything deferred here is a commitment the actor expects kept.
Let it park without ending — Disengage without closing
Keep the state saved, show that it is saved, and keep resume within reach. On return, pick up with the history, and confirm any assumptions that may have gone stale, including the capability boundary.
Connections
Calls Calibrating reliance
At Carrying the business, step 1, via Citation
the evidence attached to a direct answer, such as source, time, and uncertainty, is built in that sequence for each surface
Hands off to Delegating work
At Carrying the business, step 2, via Prompt
a framed request the agent will go away and work on becomes that sequence's prompt
Hands off to Governing autonomy
At Checking before the act, step 1, via Approval gate
the preview is what an approval acts on, and keeping the work to what was approved is that sequence's step
Hands off to Making changes, making changes safe, step 3
At Designing for trouble, step 2, via Undo
changes an aborted activity already made are undone through that sequence's safety steps
Hands off to Learning
At Setting the footing, step 1, via Onboarding
the first statement of what the agent can do is part of the actor's first encounter, which that sequence designs
Hands off to Memory and continuity
both endings leave something behind, commitments deferred at the close or state parked without one, which that sequence keeps across sessions