Skip to content

How AgentQ Works

AgentQ is an agent that works on your own Qualytics workspace. It is not a general-purpose chatbot with your data pasted into it. Every answer it gives is built by calling the same platform functions the web app calls, reading the result, and writing it up in plain language.

AgentQ uses your configured AI provider and the shared Qualytics tool catalog to help you explore datastores, build transformations, create quality checks, investigate anomalies, and take other supported actions through your Qualytics permissions. External AI clients can use the same catalog through the separate Qualytics MCP server.

AgentQ supports your data quality work, but it does not replace your judgment. Review generated interpretations and platform changes before relying on them, especially when they affect business-critical or regulated data.

Where AgentQ Appears

AgentQ shows up in three places, all backed by the same configuration and the same conversation history.

Surface What it is
Full-page chat Opens from AgentQ in the left sidebar. It carries the conversation sidebar, with the Chats and Archived tabs, and is the only surface where you can rename, archive, restore, delete, or search chats.
Floating chat The AgentQ mascot in the bottom-right corner of every page, also toggled with Q. The mascot can be dragged anywhere or tucked away at the bottom edge. It picks up the page you are on and uses it as context.
Embedded assistant The assistant inside the quality check and check template dialogs. Its chat belongs to the item once the item is saved.

How a Turn Works

A turn is one prompt and everything AgentQ does in response.

  1. You send a prompt. AgentQ receives it together with the Business Context your administrator wrote and the earlier messages in the chat.
  2. The model decides which platform tools it needs. It may call several, and it may call one, read the result, and then call another.
  3. Each call appears in the chat as a step label while it runs, so you can see what AgentQ is doing rather than watching a blank space.
  4. The model writes the answer from what the tools returned. Stop interrupts the turn, and whatever was produced is kept as a partial response.

Because the answer is assembled from live tool results, AgentQ reports what your workspace holds at that moment rather than what the model was trained on.

One Configuration for the Whole Deployment

The AI provider configuration is a single, deployment-wide setting rather than a per-user choice. A Manager or Admin selects the provider and completes any credentials it requires once, and AgentQ then becomes available to everyone with the Member role or higher. Only one provider configuration is active at a time.

Before an AI provider is connected, the AgentQ page shows a live preview of what it does, headed Ask about your data. Act on the answer., with Easy setup and Custom setup cards for Managers and Admins and a note for everyone else that an administrator connects a provider once and it turns on for all. AgentQ is set up and managed from that page, under Manage in its sidebar. See Add Integration.

The Business Context written during setup travels with every prompt, from every user, which is what makes the answers sound like they come from inside your organization. Do not put secrets or anything your provider should not process into it.

What AgentQ Is Allowed to See

Two limits apply at the same time, and the narrower one wins.

  • The access level decides what kind of data may reach the model at all, from static guidance only, through metadata, up to values from your source data. It is set once for the deployment. See AgentQ Access Controls.
  • Your own access decides which of your workspace's assets AgentQ may read on your behalf. AgentQ runs its tools as you, so a user restricted to a team sees only that team's datastores through AgentQ, exactly as they would in the web app.

When a question needs a tool the current level does not allow, AgentQ says so in the answer and explains what is switched off, rather than failing.

The Topic Guardrail

AgentQ includes an automatic guardrail that declines requests unrelated to data quality, governance, or the Qualytics platform.

What Is Allowed

Any request related to:

  • Data quality, governance, management, or engineering
  • Databases, datastores, tables, fields, schemas, or profiling
  • Anomalies, quality checks, validation, or observability
  • Transformations, ETL, computed datasets, or SQL
  • Trends, metrics, insights, or quality scores
  • Qualytics platform features and operations
  • General greetings or help requests about AgentQ itself

What Gets Blocked

Requests clearly outside the data quality domain, for example asking AgentQ to write creative content, answer general knowledge questions, or perform tasks unrelated to your data infrastructure.

How Follow-Ups Are Handled

The guardrail includes up to 2,000 characters of prior conversation context when evaluating a new message. This means short follow-ups like "yes", "that one", or "the second datastore" are automatically recognized as continuations of an on-topic conversation and pass through without being blocked.

Built-In Intelligence

  • Smart Suggestions: when you start a new conversation, AgentQ generates personalized prompt suggestions based on your data assets and active anomalies.
  • Context Awareness: AgentQ automatically detects the page you're on and injects relevant context (datastore, container, field, check, or anomaly) into the conversation.
  • Review-Before-Save Proposals: In quality check and check template dialogs, AgentQ can propose settings from a plain-language request. Nothing is saved until you apply the proposal to the form, review the highlighted changes, and save the item. Only the most recent proposal in a conversation can be applied, so an outdated configuration cannot reach the form by mistake.
  • Permission-Aware Actions: AgentQ uses your Qualytics identity and permissions, so it can only access data and perform actions available to your account.
  • Co-Authorship Tracking: when a change is made with AgentQ's help, the item's own Timeline records the assistant as a co-author alongside you, providing a record of AI-assisted actions.
  • Admin Audit Reporting: administrators can review workspace-wide AgentQ prompts, responses, tool calls, recorded actions, usage, and cost estimates for a selected date range and set of invoking users.
  • Conversation Memory: long conversations are automatically managed through context compression, keeping recent messages in full detail while summarizing older context. See Conversations, Responses & Context for details.
  • Background Streaming: if you navigate away while AgentQ is generating, the stream continues in the background. The response is waiting when you return to that session.
  • PDF Export: export any AgentQ response as a PDF for sharing and reporting.

Description Suggestions

The sparkle control next to any description editor lets you request an AI-generated description built from the asset's metadata. It is only visible when an AI provider is configured under AgentQ.

Click the sparkle, review the suggestion in the preview area, and either accept it (Use this), regenerate for a different take, or dismiss it. When the description field already holds text you have not saved, the accept button reads Replace draft instead, so it is clear what the suggestion will overwrite. The suggested text only enters the description field when you accept it. Regenerating, dismissing, or navigating away leaves the current description untouched.

Where the sparkle appears

Where What the suggestion is based on
Datastores The datastore's connection type, the containers it exposes, and its tags.
Tables The table's fields, sample field types, and any tags on the table.
Computed Tables The source containers or SQL definition, the resulting fields, and any tags on the computed table.
Files The file's fields and any tags on the file.
Computed Files The source containers or SQL definition, the resulting fields, and any tags on the computed file.
Fields The field's data type, profile statistics, and neighboring fields in the same table.
Computed Fields The transformation type, source fields, target type or SQL expression, and the parent container's field schema.
Computed Joins The source containers, join keys, the resulting fields, and any tags on the computed join.
Anomalies The anomaly's rule type, the fields it flagged, and the messages from failed checks.
Quality Checks The rule type, target fields, filter, and coverage settings.
Quality Check Templates The template's rule type and default settings.

The suggestion draws on the metadata Qualytics already stores about the asset. It does not read live data rows.

What to expect

  • Length. Suggestions are usually a sentence or two. They can run longer, and use headings and lists, when the asset has enough detail to justify it. The upper bound is the same 8,000-character limit that applies to any description you type yourself, so an accepted suggestion always fits the field.
  • Same across surfaces. The interaction (sparkle to preview to Use this / Regenerate / Dismiss) is identical whether you are on a datastore, a field, or a quality check template.
  • Language. The suggestion follows the language behavior of the AI provider and model you have configured. If you need to translate the result, edit the accepted text after using it.
  • Retries. Each Regenerate click starts a new AI request and can use a different framing, such as the audience, business process, downstream impact, or consequence of missing data. Results vary by provider and model, so two attempts may still produce similar suggestions. Usage is subject to the provider and commercial terms configured for your deployment.

When suggestions are not available

Even with an AI provider configured, the platform can still refuse a suggestion in a few cases:

Situation What the platform does
AI provider not configured The sparkle is hidden. See Add Integration to connect a provider.
Not enough context Very sparse assets (for example, an empty datastore with no containers yet) do not have enough metadata to produce a useful description. Add or connect data first, then try again.
Too much context Very large assets (for example, a container with hundreds of fields) can exceed the request budget. Narrow the scope by suggesting on a smaller asset (a field instead of the whole table, for example).
Provider timeout If the AI provider takes too long, the platform returns a timeout error. Retry, or use a faster model in your AgentQ provider configuration.
Provider error If the provider itself rejects the request (rate limit, quota, invalid API key), the error message from the provider is surfaced verbatim. See Update Integration to change model or credentials.

On every surface (Computed Tables, Computed Files, Computed Fields, Computed Joins, Quality Checks, and Anomalies), the sparkle uses the same permission as editing the description manually: if you can edit, you can also request a suggestion. On an anomaly, that means the Author team permission. Archived anomalies never show the sparkle.

What Gets Recorded

Every chat is stored with its messages, so you can reopen it later and so administrators can review it. Prompts, responses, token usage, cost estimates, tool calls, and any platform changes AgentQ made are all available to Admins in the audit. See AgentQ Audit.

Archiving a chat keeps its record. Deleting one removes it, and its turns leave the audit with it.

See Also