Patterns
Agentic navigation
The stage before any single conversation: how people get to the agent and move between agents. One bar follows whatever is on screen and offers the next step. One deck lets a person with several agents, across several parts of their life, find the right one in a swipe.
- Contextual taskbar: Use when the agent can help on many different screens and users should not have to open a chat to find out how.
- Spaces and agents: Use when one person runs several agents across separate parts of their life or work, and switching between them has to take one gesture.
Conversation core
The first thing you build: a place to ask and a place to read the answer. Every AI product ships these surfaces, so users compare yours with the best they have used. The work here is in the details: what the composer does mid-answer, how a reply arrives, what you can do with it afterwards.
- Prompt input: Use when people start work by typing a request, and need to attach a file, pick a model or stop an answer from the same place.
- Message thread: Use when users will copy, rate or regenerate an answer, and need earlier versions to stay within reach.
- Streaming answer: Use when an answer takes a few seconds to arrive and users need to see progress, where it came from and what to ask next.
- Agent chat: Use when the agent lives in a side panel next to the product and people ask it questions while they work.
- Code block: Use when the agent writes code or config the user will read, check and copy.
- Model & context: Use when users choose between models and a long conversation can run out of room without them noticing.
Trust & transparency
The stage where the answer has arrived and the user decides whether to believe it. These patterns show how the agent got there, what it read, and how sure it is. They matter most once people start acting on what the agent says.
- Thinking: Use when the agent works for more than a couple of seconds before answering and the wait needs to look like progress.
- Inline citations: Use when an answer makes factual claims from several sources and readers need to check one without leaving the page.
- Context sources: Use when the agent answers from your own documents and data, and users need to see exactly which passages it pulled.
- Confidence states: Use when the agent is sometimes unsure and a confident-sounding wrong answer would cost the user something.
- Recommendation: Use when the agent has spotted something and has a specific next step to propose, and a person makes the call.
- Feedback capture: Use when you need to learn why an answer missed, not only that it did.
Agentic control
The stage where the agent stops answering and starts doing. It covers the whole handover: agreeing what the agent may do, watching it work, stepping in, and checking what changed afterwards. Reach for these as soon as an action touches a customer, money or a record.
- Plan preview: Use when the agent is about to take several steps the user has not seen yet.
- Approval card: Use when the agent has stopped at a fork and needs the user to pick one path before it continues.
- Autonomy levels: Use when people need to set how far the agent can go on a task, from watching to acting alone.
- Permission scope: Use when people need to see and change which tools, data and spending limits the agent has.
- Task queue: Use when the agent has a backlog of tasks and people want to see and change what it does next.
- Agent status: Use when the agent runs for more than a few seconds and people need to know what it is doing and how to stop or steer it.
- Tool calls: Use when the agent edits files, runs commands or reads documents, and people want to follow each action as it happens.
- Task rows: Use when the agent runs several tasks at once and people need the status of each at a glance.
- Handoff: Use when the agent reaches the edge of what it is allowed to do and a person has to take the work or widen the limit.
- Action receipt: Use when the agent has just changed something real and people need proof of what changed, plus a short window to undo it.
- Checkpoints: Use when a long agent session changes a lot and people want a safe point to roll back to.
- Audit log: Use when someone needs to look back over everything the agent did and find the actions that need a second look.
- Error repair: Use when the agent got something wrong and has to own it, show the fix, and give people a way to take it further.
Output & generative UI
The stage where the response is a thing, not a paragraph: a document, a set of edits, a record, a chart. These patterns give that output a container the user can review, accept in part, and come back to later.
- Artifact: Use when the agent produces a document people will read, revise and take elsewhere, not a reply to scroll past.
- Diff view: Use when the agent proposes edits to code or config and people need to accept or reject each change on its own.
- Diff table: Use when the agent wants to change several records at once and people need to see every edit before it lands.
- Structured data: Use when the agent returns data in a schema and people need to read it without parsing JSON.
- Insight cards: Use when the agent has found a few things worth knowing in the data and each one needs a number and a chart to back it.
- Comparison: Use when people need to judge two models or two drafts against the same prompt and record which is better.
Ambient & beyond chat
The stage where the agent works while nobody is chatting with it. Boards, nudges, digests and inline help carry its work into the rest of the product, so people can see where things stand without reading a transcript.
- Taskboard: Use when the agent works through many tasks in the background and people should only step in where a decision is needed.
- Inline assist: Use when the agent can finish what someone is typing, in the place they are typing it.
- Nudge: Use when the agent notices something worth acting on and the user has not asked.
- Digest: Use when people come back after the agent has worked without them and need to catch up in a minute.
- Notification center: Use when the agent produces events over time and people need one place to see what is new.
- Command search: Use when people need one place to jump to anything or ask the agent, without learning the navigation.
- Agent setup: Use when someone is creating a new agent and has to give it a job, limits and tools before it goes live.
Halaska UI · Components · Patterns · Screens · Install