Agent interfaces

Show the task, not only the reply

A conversation needs more than a composer and message bubbles when work can continue outside the reply. Keep the current task, its status, and any required decision together. For multiple tasks, use a queue with a focused detail view. The person should be able to leave one task, inspect another, and return without losing a draft or the state of a pending approval.

The AI chat study provides conversations, local response templates, an inspector, and Markdown export. The Agents study explores queues and review. Both use sample or local behavior. Start with one of these interaction patterns, then connect your own provider through an adapter with explicit loading, completion, cancellation, and error states.

Separate proposed work from execution

The host application must enforce permissions and tool boundaries. A confirmation control makes a decision visible; it is not authorization by itself. Keep each stage inspectable:

• Proposed: describe the exact action, target, and relevant inputs before asking for approval.
• Awaiting review: keep approve and reject available, and invalidate approval if the action changes.
• Running: identify the action in progress and provide cancellation when the host supports it.
• Completed: report the result returned by the host, with a link or identifier where available.
• Failed or interrupted: preserve context and explain whether retry could repeat an external effect.

Keep output usable

People should be able to select, copy, inspect, and export useful output. Put source identity and update time beside retrieved records. Use a widget for structured content such as a schedule or a comparison, with the same loading, empty, and error behavior used elsewhere in the product.

Streaming text needs a stable reading position and a way to stop. Do not move keyboard focus when new content arrives. Announce meaningful state changes without reading every token aloud. Treat generated markup as untrusted content and choose rendering boundaries in the host application.

Vlak's agent documentation serves a different purpose: it helps a coding agent build with the library. It provides the component reference, tokens, registry, and MCP tools. That does not turn a product prototype into a connected assistant; the application still supplies its models, authentication, tools, and storage.