Start a conversation
Requirements, execution steps, and result definitions are often scattered between the requester and the executor, which causes most rework. A conversation brings questions, execution, and verification together so that you can expand each tool call and data source.
Core example
Click New task in the sidebar, attach your data, and describe the goal directly:
Count the order volume and period-over-period change in each region over the last 7 days, sort the results by order volume in descending order, and identify the region with the largest fluctuation.The reply provides conclusions and supporting data. Tool calls appear as cards that you can expand to view the query that ran and the number of returned rows. When you ask follow-up questions, the context continues automatically, so you do not need to repeat the background.
To assign a specific agent to a task, see Agent management. To run a task on a fixed schedule, see Scheduled tasks.
Scenarios
The web workbench is the full-featured primary entry point. You manage conversations, agents, skills, knowledge bases, scheduled tasks, connectors, and artifacts there. IM channels support conversations and result delivery, but you must return to the web workbench to create and configure assets.
Before users can use an IM channel, an administrator must register the channel and associate it with an agent, and members must pair their accounts. See IM channel management.
Start a conversational task in the web workbench
A complete task includes describing the goal, confirming the execution plan, observing execution, and verifying the results. The confirmation step determines whether later execution deviates from your expectations, so do not skip it.
Start a new conversation. Click New task in the workbench sidebar. Keep each conversation focused on one topic because unrelated questions interfere with each other’s context and reduce the accuracy of follow-up questions.
Describe the result that you want to achieve instead of the execution steps. State the goal, available materials, and expected output. For example, say: “Based on this requirements document, create a launch checklist and identify items that require manual confirmation.” A more specific goal produces an execution plan with less deviation.
Review the execution plan. Confirm execution only after Ada provides the Execution plan. Verify the data scope, breakdown dimensions, and output format. If they do not meet your expectations, add constraints directly instead of starting another conversation.
Observe execution. Steps that involve external operations appear as tool call cards showing execution status, input parameters, and output results. Expand them at any time. To stop execution, click the stop button next to the input box or press
Escape.Review the results. Content-modifying operations display added and removed lines as a diff. When a knowledge base is attached, the reply includes Wiki references. Open them to confirm that the referenced definition entries and knowledge base version are the expected ones. If you upload only files, expand the tool calls to review the execution process.
When an agent requests permission during execution, choose from Allow once, Deny once, Always allow, and Always deny. If the impact scope is uncertain, choose Allow once. Before you confirm a delete, publish, overwrite, or external-send operation, verify the target, scope, and recovery method. Do not use Always allow for these operations. The labels identify option behavior; the button text displayed on the confirmation card is authoritative.
Continue context in a multi-turn conversation
Ask a follow-up question as soon as a reply finishes. The agent continues working with the current context, so you do not need to repeat the background. To specify the executor, type @ in the input box, search for and select an agent, and bypass Ada’s assignment step.
When an answer is not organized as expected, regenerate it. Regeneration creates a branch in the current conversation. Use branch navigation to switch between versions. Each turn retains up to 5 branches, and earlier answers are not lost.
After a conversation exceeds 5 turns, use the message navigation on the right to jump quickly between messages. With the cursor at the beginning or end of the input box, press ↑ and ↓ to browse recently sent prompts.
Add context materials to a task
Without sufficient background material, an agent can only respond based on general knowledge. Use Add content on the left side of the input box to add any of the following four material types, and combine them as needed.
Upload files for one-time document and spreadsheet input. Image attachments are also supported, and you can remove them in the preview area before sending. Attach knowledge bases when your task needs consistent definitions. Select one or more knowledge bases for the next message; attached knowledge bases appear as tags that you can remove individually. See Knowledge base management. Attach skills or connectors to add execution methods and data channels to the current task when data is connected but not selected by default. Paste or upload images to interpret screenshots, charts, and interface content. Speech input is also supported. Click the microphone button below the input box to start recording, then click it again to stop and transcribe. The transcription is placed directly in the input box. On first use, allow microphone access in the browser. If the interface reports that the microphone is unavailable, grant access and try again.
Your selection of materials directly affects the execution path. When you provide only files, the agent answers within the files’ scope. The agent retrieves data from the engine only after you attach a connector.
Review and retain execution results
After you receive a reply, you can do four things. To review how a result was calculated, expand the collapsed area above the reply. Its title is Processing while execution runs and Processed after it finishes. To take results with you, preview web page and spreadsheet outputs online from their cards, then download spreadsheets from their cards. To confirm changes, review previews and diffs for modified files in the review panel on the right. To confirm the definitions that were used, expand the Wiki references list in a reply that has an attached knowledge base. The body renders as Markdown, so code, formulas, and flowcharts display directly without additional processing.
You can copy or regenerate any reply. When troubleshooting, use Copy troubleshooting information to include the conversation and request identifiers and locate that execution.
Conversations remain in the history list, where you can search, rename, and delete them. You can manage scheduled tasks from Scheduled tasks in the left-side menu or from the Scheduled tab in the conversation list area. Long conversations load history in two stages. For messages already loaded in the current conversation, scroll to the top to expand an earlier section automatically, or click Show earlier messages to expand it manually. After all messages are expanded, Load earlier messages appears at the top of the message area. Click it to retrieve older history from the server. Loading… appears while loading. If loading fails, Failed to load earlier messages. Try again. appears, and you can click again. If you see The current history cannot be loaded. Re-enter the conversation to restore it., re-enter that conversation. A proven conversation workflow can be turned into a scheduled task that runs periodically. See Scheduled tasks. For content that needs long-term retention and sharing across members, have the agent push output as an artifact or import it manually on the artifact page. See Artifact management.
Start tasks and receive results in IM channels
IM channels suit two scenarios: asking questions directly in a chat window and receiving results and alerts from scheduled tasks. Team members can obtain conclusions without entering the web workbench.
Send a question directly in an IM conversation that is associated with an agent. The agent replies in the chat window, and the conversation is synchronized to the Channel tab in the web workbench history area. Return to the web workbench to view the complete execution process and tool call records. Scheduled task results can be delivered through an IM channel to paired recipients. Set the delivery target in the push configuration of the task form. See Scheduled tasks.
IM interactions are limited by the chat window. Use the web workbench for complex multi-turn exploration, material uploads, and asset configuration.
Quotas and limits
Limit | Description |
Conversation branches | Each turn retains up to 5 branches. |
Message navigation | Available after a conversation exceeds 5 turns. |
Conversation attachments | Up to 6 files per upload, with a combined size of no more than 48 MB. |
Knowledge base attachment | A conversation can attach one or more knowledge bases, which you can remove individually. |
IM channel capabilities | IM channels provide conversations and result delivery. Create and configure assets in the web workbench. |
FAQ
Q: Are earlier answers still available after regeneration?
A: Yes. Regeneration creates a branch in the current conversation. Use branch navigation to switch between historical versions, and the original answer remains intact.
Q: The tool probe passed. Why does the call still fail?
A: A probe verifies only the capability handshake. It does not cover authorization or parameters. Check whether permission confirmation was approved, tool parameters are complete, and the external service returned an actual business result instead of an empty response. For data tools, also confirm that the data access policy did not block the current user.
Q: Why can I not see the agent or skill that I need?
A: Clear the search criteria first, then check the ownership filter in the list. If the resource is still unavailable, contact the resource maintainer to confirm that the resource was created and its ownership covers the current member.
Q: What is the Channel tab in conversation history?
A: This tab groups conversations from IM channels separately from conversations started in the web workbench. You can view the complete execution process for IM conversations, including tool call records that were not expanded in the chat window.