Give the Agent a task
A Runtime Agent works inside an app. It can answer questions, use the tools enabled for that app, and work through several steps. State the desired result and provide the relevant material. For example: “Compare these two proposals and create a short summary of the differences.”
The app decides which tools, skills, and controls are available. An Agent cannot use a capability just because another app has it.
Creating your own app? You can ask Builder to add the ready-made Agent Conversation component or Chat Widget. This page explains how to use the conversation once it is in an app.
Manage conversations
A Session is a saved conversation. Open the Sessions panel to start a new one or find earlier work. Search, rename, or pin a session when you need to keep it accessible. Continue an existing session when its history is relevant; start a new one for an unrelated task.
Deleting a session is different from starting a new conversation. Review the deletion action if you need its history or generated files later.
Attach files and context
Use the attachment control or drag a supported file into the conversation. You can also add selected text as context and use voice dictation where transcription is configured.
Attachments and file references are attached to the message you send. Review the pending items beside the composer and remove anything you do not intend to include. They are not a permanent instruction to include the same files in every future message.
An app may provide Session Context, such as the current trip or document. The app owns this context. A local file reference points to a file on the host; an upload from a remote device sends file contents to that host.
Follow and steer work
Tool activity shows what the Agent is attempting and whether it completed or failed. A request for approval pauses the affected action until it is resolved. Read the requested access before deciding; see approval states.
- Steer
- Add guidance for the active task.
- Follow-up
- Queue a request to continue after the current work.
- Stop
- Interrupt the active work. Files already written or external actions already completed are not rolled back.
After a failure or interruption, inspect the result before asking the Agent to repeat an action. Work in a different session is separate.
Respond to choices and drafts
An Agent may show a choice card when it needs a decision. Choose an option, submit the selected options on a multi-choice card, or type a response. Answered cards stay in history.
A Draft contains text you can copy or discuss. Comment adds that draft to the next message's context; it does not send a message or change the original draft by itself. Type the changes you want and send them.
Answers can include Markdown, tables, formulas, diagrams, sources, and images. Check the cited material and any calculations before relying on the result.
Preview and use generated files
An Artifact is a file published as a session result. Open its preview to inspect it. Supported document, image, and other file previews appear in the conversation's preview area; image previews support closer inspection with zoom and pan.
- Find the file card in the conversation, or open Show files and look under Artifacts in the Context panel.
- Select the file name or its Preview control to open it.
- Use Comment to request changes, or check the file's More menu for available actions.
Use Comment on a file card to refer to that exact artifact in your next message. Use the available Copy, Share, Open, or Download actions to take the file into another workflow. Actions vary by file type and device; local Electron file copying and mobile sharing are not available in every browser.
For an example, see how to preview and revise a PDF in One Shot.
Continue long conversations and team work
For long conversations, AppJuice can summarize older context before continuing model requests. Saved conversation history and the context sent to a model are different: a model does not necessarily receive every earlier message verbatim. Restate an important detail if it is missing from a later answer.
An app may coordinate multiple Agent roles and show their work in a collaboration view. The app defines the order, result sharing, and failure handling. This is an app feature, rather than a team mode available in every conversation.