Agent Research

Agent Research, labeled Agentic Research in the dashboard, plans a task, inspects your database schema, reads records, and runs calculations in an isolated environment. With the relevant permissions, it can also update data and schemas, create dashboards and visualizations, save documents and scripts, and perform database maintenance.

Use it for comparisons, trends, statistical questions, and tasks that turn findings into resources. For quick lookups and conversational exploration, use the AI Chat workspace.

1. Sign in to Onyx Cloud and select an organization and one of its databases.
2. Open Agentic Research in the navigation, or visit /research.
3. Enter a question. Name the population, date range, comparison, and outcome you want to understand. Mention relevant tables or fields if you know them.
4. Choose the available account source and Model, set a Thinking level when supported, and review Permissions.
5. Select Start research. The task appears in your research history and begins reporting progress.

The workspace shows which models are available for your environment. If research is unavailable, its status message explains why; starting a task stays disabled until it is enabled.

Without saved preferences, tasks start with read-only access. Open Permissions to grant individual actions, choose Read-only, or select Full Control for all actions your current role allows. Full Control remains scoped to the selected organization and database.

• Database: schema changes, data mutations, snapshots, compression, database deployments, and re-indexing.
• Workspace resources: visualizations, dashboards, documents, and draft organization agents.
• Scripts: create saved scripts, execute them, and configure recurring schedules.

The dashboard remembers your permission selections for new tasks and follow-ups in this database. Review them before starting. Your role is checked during execution, and unavailable actions cannot be granted.

Script execution runs an authorized saved script in the research environment. Writes still need the data mutation grant. Scheduling requires script schedules, script execution, and data mutation permissions. Enabled schedules continue after a research task ends or is canceled; manage them in the Scripts scheduler.

Use the models available through Onyx or connect your personal ChatGPT/Codex account through the account controls. The model list follows the selected source and its available capabilities.

Choose a Thinking level from the options supported by that model. Higher levels may take longer. The task records the model, account source, and thinking level so you can review how it ran.

Usage is shown when reported by the provider. Connected accounts remain subject to their provider’s access and usage limits. Reconnect an unavailable account or choose another available source before starting new work.

The agent can page through all relevant records and process them in batches. Page size controls each batch, not the total number of records it can investigate. If the database rejects a query with a cardinality error, the agent can adjust the query and continue.

Research does not impose task budgets on query rows, steps, duration, or model spending. Provider limits still apply. Use Cancel research whenever you want to stop.

The activity feed explains what the agent is investigating, what it has learned, and what it will check next. Expand Progress timeline to revisit those updates. The task header shows elapsed time and model usage.

• Queued / In progress: the task is preparing or investigating. Select Cancel research to stop an active task.
• Completed: the answer is ready to review.
• Failed / Cancelled: review the error or final progress event before starting another investigation.

You can leave the page and return through Research history. Refreshing the dashboard does not restart the task. When saved work is available, select Resumeto continue an interrupted task. Use Follow up to investigate a related question.

The Answer gives the conclusion in a short paragraph or a few brief bullets, with the numbers and caveats that matter. Open Details and evidence when you want the supporting analysis.

Check the population, sample sizes, assumptions, and missing data before using a result. If the available data cannot support a conclusion, the answer should say what is missing.

Downloads appear under Supporting files. Use the download button to save available datasets, analysis code, charts, or documents. The files produced depend on the investigation. Reports and uploaded artifacts remain available in your task history after the analysis workspace closes.

Research requires organization membership and access to the selected database. Task history, reports, and downloads are scoped to the user who created the task, the organization, and the database.

The agent can read physical table columns and follow continuation pages. Schema changes and data mutations require their corresponding permission toggles and your database role. Research does not execute relationship resolvers. Saved-script execution needs its own grant; calculations and script execution run in the isolated research environment. Access is checked throughout the task.

• No active database: select an organization and database in the dashboard before starting.
• No models available or research disabled: follow the workspace status message and ask your administrator to check research availability.
• Failed task: inspect the error and last update. Resume saved work after resolving the issue, or start a follow-up.
• Missing history: check that you are signed in as the task creator with the same organization and database selected.