Skip to main content

Agents

Agents turn a clear business task into a scheduled or triggered cloud run. Define the records to inspect, the result you expect and which actions need review. A saved schedule is configuration; it is not evidence that a run or its proposed changes completed.

This review created Weekly bookkeeping rehearsal, kept it Paused, and confirmed its name, instructions and credit limits after a reload. It never ran. All five accounting-template dialogs were inspected without carrying out their accounting actions.

Phone receptionist​

Use the Phone receptionist template to prepare an agent that answers business calls. Its Phone section has Answers phone calls, greeting, voice, language, tools, customer-record access and call-length settings. Assign the saved agent to a line under Assets → Phone numbers, or Phone → AI Proxy for a forwarded line.

Phone answering is separate from scheduled work: pausing the schedule does not stop the agent answering its assigned lines. Disable Answers phone calls or change the line's routing to stop that behavior. See AI agents on the phone for tools and privacy settings, and Phone for call summaries, recordings and follow-up.

Cloud Agents​

Open Agents in the intended business. Choose New Agent or a template, then review the instructions and trigger before saving.

  1. Give the agent a recognizable name.
  2. Write an observable result: the date range, records to inspect, missing-information handling and intended output.
  3. Choose a schedule or supported event, watch or reminder trigger. Confirm the time, relevant days and timezone.
  4. Review the enabled state. Use Paused when preparing a task that should not trigger yet.
  5. Set credit limits where appropriate and inspect any optional role/style instructions.
  6. Create the agent, reload the page and reopen Edit to check the saved settings.
Actual Weekly bookkeeping rehearsal setup showing Paused, Monday eight AM, a read-only review brief and 100 per run / 500 per month credit limits
The filled configuration before Create agent. The paused task was subsequently saved and its exact instructions and limits reopened after a full reload. It was never enabled or run.
Actual Weekly bookkeeping rehearsal setup showing Paused, Monday eight AM, a read-only review brief and 100 per run / 500 per month credit limits

The filled configuration before Create agent. The paused task was subsequently saved and its exact instructions and limits reopened after a full reload. It was never enabled or run.

The rehearsal asks for a summary of uncertain bank transactions and recurring payees. It explicitly excludes categorization, matching, rule creation, messages and record changes. Its schedule is Monday at 8:00 AM, with 100 credits per run and 500 per month. These are example limits, not a forecast of the credits a bookkeeping run needs.

Choose the right trigger​

TriggerWhat to check
ScheduleDaily, weekdays, weekly, monthly or a supported one-time date; confirm timezone.
Customer inquiryIntended email, chat or completed-call sources and their connections.
Email or callWhich email category or completed-call context the task receives.
Form or work orderThe submitted form, newly created job or completed-job event.
App eventThe app and event name, rather than every event from every app.
WatchThe supported condition, such as overdue invoices or low stock, and check interval.
ReminderThe intended future date and task instructions.

A trigger does not connect its source channel. Verify the related inbox, app, form or operational record separately. Event agents do not become ordinary clock schedules just because their editor also displays scheduling controls.

Use Run now only when you intend to start work immediately. An agent can continue in the cloud after the tab closes. Pausing future triggers and stopping an active run are separate actions. None of these run or stop behaviors was exercised for the rehearsal.

Agent Budgets​

Credit limits exist at the individual-agent and broader settings levels. Review both before starting paid work.

In an agent's editor, expand Credit limits and set Per run and Per month. Empty means no limit at that level. Agents → Settings → Credit limits shows broader usage and limits; inspect the selected business and current settings before changing them.

The implementation checks applicable limits before reserving credits for paid work. A tighter applicable run limit can constrain a looser one, and both overall and individual-agent monthly limits matter. Usage distinguishes spent credits from credits reserved for work in progress.

Monthly limits reset on the first day of the month in UTC. A request already running can finish above its estimate, so a configured limit is not a guarantee of an exact final charge. Read a budget-related pause before deciding whether to adjust a limit and resume.

The rehearsal's 100 per run / 500 per month limits survived a reload. Broader settings were inspected without saving changes. No paid run, reservation, budget pause, monthly reset or resumed execution was tested. Do not interpret the screenshot as a completed budget-enforcement test.

Agent Checklists​

A cloud run can display Task progress with individual items marked Next, Working, Blocked or Done, plus a completed count. This helps distinguish an unfinished step from a run-level status.

Open the relevant run conversation and inspect the item details alongside the actual output and proposed actions. A checklist item marked Done does not by itself prove a payment, publication or external message completed. Check the corresponding record or delivery result.

If a step is Blocked, read what is missing before resuming. A question waiting for your answer, a credit pause and a failed external service need different responses. Check that the next instruction addresses the actual blocker.

Checklist rendering and durable task-state handling were confirmed in source. The paused rehearsal had no run, so no populated checklist, progress transition or completion result was produced for its screenshots. This guide does not fabricate a run merely to show progress.

Accounting templates​

Choose Money in the template filters to find the accounting starters. Each template is an editable brief and default schedule, not an autonomous guarantee that the books will be correct.

Actual Money templates including Bookkeeper, Invoice chaser, Bills due, Payroll prep and Sales tax watch
The real template chooser. All five relevant setup dialogs were opened and their instructions inspected; no accounting template was run.
Actual Money templates including Bookkeeper, Invoice chaser, Bills due, Payroll prep and Sales tax watch

The real template chooser. All five relevant setup dialogs were opened and their instructions inspected; no accounting template was run.

Before using these starters, make sure the underlying Books, Banking, Invoices, Bills, Payroll and Sales tax records are current. A missing integration or incomplete record can limit the result even when the template itself opens.

Bookkeeper Agent​

The Bookkeeper template starts with a weekly review of bank lines, categorization/matching and recurring payees. Its default is Monday at 8:00 AM.

For a plumbing workshop, use it to prepare a review of supplier purchases, incoming customer payments and uncertain bank lines. Before enabling it:

  1. Inspect the bank accounts and transactions that should be included.
  2. Open Money → Bookkeeper and read the complete brief.
  3. Decide whether you want proposed categorization, a summary only or some other narrower result. Rewrite the instructions accordingly.
  4. Define how uncertain lines should be handled and which rules require your decision.
  5. Review the schedule, pause state and limits, then save and reopen the task.

The shipped brief proposes rules for recurring payees and asks the agent to list uncertain items with a question. The demonstrated rehearsal deliberately narrowed this to a read-only summary. Neither the shipped brief's categorization/matching behavior nor an accounting result was executed here.

Invoice Follow-up Agent​

The UI template is named Invoice chaser, with a default weekday 9:30 AM schedule. It starts from receivables aging and proposes polite reminders for invoices more than seven days overdue when no reminder was sent in the preceding seven days. Its brief becomes firmer after 30 days while excluding new late fees or threats.

Open the template and review its overdue threshold, reminder interval and wording against your business's collection process. Confirm the invoice balance, customer and earlier contact before approving a reminder. A paid, disputed or adjusted invoice needs different handling from an ordinary overdue balance.

Ask for a summary of the total overdue, proposed reminders and largest balances so the next decision is visible. The template dialog and its instructions were inspected. No reminder was created, approved, sent or delivered, and no customer was contacted during this review.

Bills Due Agent​

The Bills due template defaults to Monday at 8:30 AM. Its brief reviews overdue bills and those due in the next seven days, compares them with available cash and flags possible duplicates with the same vendor and amount within 30 days.

For a weekly supplier-payment review, check that bills have accurate vendors, due dates and balances. Ask the agent to explain which amounts are due, which records look duplicated and what remains uncertain. A duplicate warning is a reason to inspect records, not permission to remove a bill automatically.

The starter only proposes payment for bills you have explicitly instructed it to pay on schedule; it asks for the rest. Review the actual payment proposal and funding details before approving. No payment proposal, bank transfer, bill settlement or duplicate deletion was performed here.

Payroll Prep Agent​

The Payroll prep template defaults to Thursday at 10:00 AM. It checks whether the next pay period ends within three days or an ended period still lacks a run, then inspects timesheet issues before proposing payroll preparation.

Use it to surface unapproved clock-ins, open entries or missing hours before a manager reviews payroll. Confirm the period, usual pay date and affected staff in Payroll and Shifts.

The shipped brief says to prepare the run for review and never approve it. Gross, net and included staff should be checked in Accounting → Payroll. If no run is due, the intended result is a concise next-pay-date summary.

The setup brief was inspected only. No employee hours changed, pay run was prepared/approved, payment made or payroll calculation verified by this agent review.

Sales Tax Watch Agent​

The Sales tax watch template defaults to the first day of each month at 9:00 AM. It checks returns due in the next 30 days, prepares the corresponding worksheet and highlights unusual entries for review.

Use its output as a preparation aid: inspect sales with missing tax, purchase tax that may need review, changes from the previous period, amounts and the due date. Reconcile the result with the actual Sales tax worksheet and underlying records.

The shipped brief explicitly excludes filing a return or marking it filed. A task summary saying the return is ready does not establish a filed return or completed payment. No return worksheet was generated by this agent, filing status changed, return submitted or payment made here.

Review proposed changes​

Needs you can surface pending actions, drafts and pauses. Read the affected records and values before choosing Approve, Skip, Review or Resume. A draft reminder dismissed with Later is not published by that dismissal.

In Settings → Approvals, Ask me every time and Auto-approve low risk have different effects. Selecting the latter can save the setting immediately. The implementation uses a separate review step for eligible low-risk changes; sensitive actions such as money, sending, publishing, domains and access still require review. This review did not change the approval mode or approve an existing pending action.

Permissions, runtime availability and actual action results still matter. Review the run conversation and the resulting business record instead of treating a template name or completion message as proof.

Troubleshooting​

SymptomWhat to check
Saved task does not triggerEnabled state, trigger, timezone, connection and next-run information. Paused tasks deliberately do not run on schedule.
Run stops for creditsThe displayed pause reason, applicable limits and available credits; raising a limit and resuming are separate decisions.
Task has a completed checklist but no expected recordInspect the proposed actions and actual record; progress labels do not replace execution evidence.
Reminder or payment seems readyReview the exact customer/vendor, balance, authority and delivery/payment action before approving.
Template opens but cannot use its recordsCheck underlying Books setup, integration access and data completeness.
Instructions changed but output seems oldConfirm which saved schedule and run produced that output. A new configuration does not rewrite an earlier run.

Continue with Working with AI, Agent memory, Integrations, Work orders and Calendar.

Follow agent work from AI​

The current /ai workspace brings task conversations together, including runs opened from configured agents and notifications. The agents button indicates work or decisions waiting for you. Open a run to read its question, progress and proposals; use Agents for the reusable instructions, trigger and limits.

Answer a question in the run that asked it. Pausing a schedule prevents future scheduled work; Stop the task affects an active run. Settle removes a card from the active AI canvas and is not a cancellation command. See Working with AI.