Agents run interactive tasks in your data plane with the launcher’s access, limited by the capabilities an administrator consents to. The Automations entry appears when the feature is enabled for your tenant and your role can read agents. In a tenant without the feature, the entry is hidden and opening the page directly shows that work and agents are not included.

Find and install an agent

Open Automations, follow the Agents link and choose Catalog to see registered agents that are not installed. Open an agent to review its description, version, source image, and requested capabilities. In Consent & budget, select the capabilities the installation may use and set the daily run limit, time limit, and maximum reported cost per run. Installing requires permission to manage agent installations. The Installed tab shows each installation’s version, external connections, status, and runs today. Use its filter to find installations that need attention or are parked. The Tasks tab shows tasks across agents, including Awaiting you for decisions you may make.

Use the routine runner

When automatic setup is enabled, Dippr installs the routine runner for eligible new and existing tenants with platform-policy consent for data reads and writes. The defaults are $1 per run, 20 runs per day and a five-minute timeout. Runs still use the launching person’s or subscriber’s access within that ceiling. These limits do not guarantee a daily invoice cap. Installation starts no model call or schedule. Existing installations retain their image, consent and budget. Paused runners stay paused. Removing a runner prevents automatic reinstall, including after a later manual reinstall. A runner removed before removal tracking can be installed again automatically. An authorized administrator can still install it manually. When automatic setup is disabled or the runner has been removed, the Automations page shows Set up automations to people with installation permission when the runner is missing. This opens the routine runner page for consent and installation. Other people see a message asking an admin to set it up from Agents. A paused runner also shows a notice with a link for people with permission to update installations. Incomplete consent links to the Consent & budget tab; a runner that is being removed shows a message and no link until removal finishes. A routine runner is a built-in agent that investigates your instruction with governed SQL and reports findings with supporting evidence. When your deployment enables it, Routine runner appears in the Catalog for your tenant. When automatic setup is disabled or the runner has been removed, an administrator must install it and consent to its data access before anyone starts a task. Choose an engine and make the first task message a complete instruction: name the sources, checks, thresholds, and when to stop. The runner can ask a question or propose SQL for review. Review a proposal before approving it. A proposed action does not execute merely because it appears in a result. Automations read data you can read, including your workspace databases and workspace databases shared with you. What they read can be sent to the configured model provider to investigate your instruction. Each person’s run uses that person’s current access, within installation consent and budgets. Scheduled SQL changes require the routine owner’s approval in the console; per-person routines are read-only. The model connection shows Managed by your deployment. Prompts and query results leave the data plane through this connection. Your deployment administrator can change its endpoint without reinstalling the runner or changing your data consent. A hostname grant on an older installation does not control this connection; removing that grant does not disable model access. Park or delete the installation to disable execution. Saving consent on the installation can only narrow it. Workspace-database reach was added to existing installations automatically and follows each person’s own access. Your deployment supplies the model connection. You cannot change its credential, endpoint, or model through a task or catalog publication. A routine is limited by its cost per run and time limit. A run stopped before its cost ceiling, or by a limit on turns, tokens, tool calls, or time, files a partial report saying what it checked. If a model response overshoots the per-run cost ceiling, the run fails as budget exhausted without a report. A partial report does not close earlier findings it did not re-check. A failed or incomplete report does not establish that earlier findings have been resolved. To save a schedule and track items across runs, use Automations with this installation. Starting an interactive task here does not create a schedule. The Agents page shows the agent catalog and interactive tasks. Shape ongoing work in your AI and review it under Automations.

Start and follow a task

Open an installed agent and select New task. Choose the data engine the task will use, then type your request. The dialog estimates remaining runs from tasks you can read; other people’s tasks may have used the shared limit. When your visible runs reach the limit, the button is disabled. The service checks the shared limit and engine again when you start. If no data engine is available, ask an administrator to add one before starting a task. Starting opens the task view. If prompt delivery fails after a task is created, the dialog keeps that task for a retry or cancellation. You can also open the task, check its transcript, and send the prompt there without starting another run. The task view shows a successful structured Result first: its formatted summary, findings with supporting evidence and query references, suggested next steps, and coverage limitations. The Conversation shows readable messages and accumulated agent text. When no failure banner shows the reason, event errors appear in Conversation with readable labels, such as Budget exceeded. Markdown supports headings (displayed at one heading size), paragraphs, flat unordered lists using - or *, bold text, inline code, fenced code blocks, and HTTP or HTTPS links. Headings and lists can directly follow text without a blank line; ordered lists, nested lists and italic text are not interpreted; raw HTML and images are displayed as text. Markdown tables show the first eight rows; select Show all to read the rest. Open Details for the raw events and lifecycle trace. A result from a failed, cancelled, paused, or incomplete run stays provisional in Details. Results and conversation content stay on the data plane. If the data plane cannot be reached, the trace and decisions remain visible and the page says that conversation content is unavailable. An active task updates as events arrive, with periodic refresh for lifecycle changes. Tasks in the global Tasks tab remain linked after their installation is deleted, so retained task history can still be inspected. While a task is queued or running, Cancel task sits on the task page next to Send message. When a task fails, the page shows the reason the agent reported, such as a missing model key, above the final state. When an agent asks a question or proposes a plan, send a response in the input panel. When it proposes an action held for review, the launcher or an authorized data administrator can approve, request changes, or reject it. Review the proposed effects before approval. Cancelling a task or rejecting its proposal shows the final state at once; the agent’s sandbox stops, and the task’s access is removed, shortly after. Approval stays disabled until the current proposal loads from the data plane, and the task service rejects approval if that proposal changes. Approval adds no access: the work still runs with the launcher’s access within the installation’s consent. Approving is separate from running: a connection granted only agents:run can start tasks but cannot approve their proposals, and neither can any token exchanged from an AI connection. For interactive tasks, a sign-in outside the console that explicitly requested agents:review and was granted it, such as dip login, can approve (see OAuth scopes). Scheduled writes still require their owner’s console approval. After an interactive task finishes, a follow-up reopens that task under the original launcher’s access, including when another person sends the message. Open an installed agent’s Consent & budget tab to narrow its capabilities or change its budget. Narrowing takes effect for future actions after the policy update. Widening requires consent again. The installation limits daily runs, cost per run and run time. A run stopped by its turn, token, tool-call or time limit files a partial report saying what it checked. If a model response overshoots the per-run cost ceiling, the run fails as budget exhausted without a report. A partial report does not close earlier findings it did not re-check. The Configuration and Versions tabs are read only. To change an agent definition, publish a new version from its source project.

Scheduled reports and approvals

An automation is a scheduled job with a record and a report for each run. Open Automations to read the latest report, inspect its read-only findings and review a proposed write. The Needs you strip points to approvals waiting, failed last runs and stopped automations. Only the automation owner can approve or reject its scheduled write. The review page shows the exact statement; a link to review never approves or executes it. Per-person runs keep separate access, and administrator status does not grant access to private summaries, evidence, instructions or steps. Interactive agent decisions remain in Agents → Tasks → Awaiting you. The Automations rail badge counts automations needing your attention: approvals, failed last runs and stopped automations. Open Automations to see whether the Needs you check is incomplete. The Tasks tab count shows tasks Awaiting you, the default filter. When nothing is waiting on you, the page shows how many other tasks are available. Select Show all to see them. Task results use readable evidence labels and statuses such as Needs attention. A recorded cost smaller than half a cent appears as <$0.01. On a routine run stopped by its model cost limit, the smaller recorded amount is shown. The run message names a recorded cost, text, model step, tool step, time, or data limit when available. A missing summary alone does not mean a run failed. On a scheduled task, Instruction holds the full request. Its shortened title appears at the top, and Details holds its identifier and trace. If you cannot open a task because it is private or no longer available, the page explains who can open it and offers Back to Automations. If engines are slow to load when starting a task, the agent page remains available and the dialog shows Loading data engines… followed by the engines or a load error.