Skip to main content
A scheduled agent is an Accounting Intelligence instruction that runs on a cadence without anyone in the composer. Use Agents for recurring work: a Monday list of newly overdue invoices, a daily close-status summary, an unreconciled-items report before a period lock. Scheduled agents appear under Scheduled in the Agents panel, each with a cadence badge such as Recurring.

​
Definition fields

A scheduled agent is defined by five fields, set from Scheduled Agents in the Agents panel. Two of them, entity scope and accounting basis, determine the figures every run returns. Creating, editing, pausing, resuming, deleting, and rolling back an agent are all restricted to organization admins. Anyone with Accounting Intelligence access can see the schedule list, so a non-admin can tell that an agent exists and when it runs without being able to change it. An organization can hold 50 scheduled agents at a time. There is no per-agent connector selection. An agent reaches the same connectors as the person who owns it, which are the ones that person has personally connected. See how to connect an MCP connector for why connections are per-user rather than per-organization.

​
Cadence options

Cadence is either a repeating interval or a calendar schedule, and the picker offers five modes. Every calendar mode carries its own timezone. There is no separate daily mode. For a once-a-day agent, set an interval of one day, or use Weekly with every weekday selected if you want it to skip weekends. Times resolve against the timezone stored on the schedule. Any IANA timezone is accepted, such as America/New_York. A schedule saved without one resolves against UTC, which is rarely what you want for a close-related agent. The minute you choose is adjusted by up to four minutes when the schedule is saved. DualEntry spreads scheduled runs across a five-minute window so that agents created on the same round hour do not all fire at once. An agent set for 09:00 may therefore run at 09:03.

​
Where results appear

Each scheduled agent owns one conversation for its lifetime, and that conversation is the only place its output goes. Every run appends one prompt and one response to it, so an agent named “Monday AR review” accumulates one exchange per week that you read in your chat list. There is no email digest, no Slack delivery, and no notification of any kind. Reviewing a scheduled agent means opening its conversation. An agent whose output nobody opens produces nothing that anyone reads, which is the practical limit on how many agents are worth creating. Who can open that conversation is narrower than who can see the agent. The owner always can, and an organization admin can, so that scheduled output stays auditable without anyone impersonating the owner. Everyone else needs the session shared with the organization. Check that the people who are supposed to act on an agent’s output can actually read it before you rely on the agent. Deleting an agent deletes its conversation along with it. Pause the agent instead when you want the schedule stopped and the history kept. Because the output accumulates in one place, an agent running for a year holds a year of runs in a single conversation. Scroll back to compare a current result against an earlier one, which is often faster than re-running the query for a past date.

​
Run behavior

A scheduled agent runs unattended, which changes what its permissions, its scope, and its failures mean compared to the same instruction typed into the composer. An agent runs under the permissions of the user who owns it. Results reflect that person’s role scope, which matters when the owner changes teams or leaves. Review agent ownership when roles change. Entity scope and accounting basis come from the agent definition rather than from a selector someone sets at run time. An agent left unscoped returns consolidated figures on every run, and no one is in the composer to catch it. Each run is written to the audit trail with the same detail as an interactive conversation, attributed to the owner and tagged as a scheduled run, so an unattended result is as traceable afterwards as one you asked for yourself. A failed run increments a failure counter on the agent. Three consecutive failures move the agent to an errored state and stop the schedule, which keeps a broken instruction from running every day indefinitely. The owner is not notified when this happens, so an agent you depend on is worth opening periodically to confirm it is still running. Pausing an agent stops the schedule and keeps both the definition and the history. Resuming restarts it on the same cadence and clears the failure counter. The agent detail view carries the current status, the last error, and the version history of the definition, so a result you cannot reproduce interactively can be traced to the wording and scope it ran under. The runs themselves live in the agent’s conversation.

​
Read and write boundaries

A scheduled agent reads and reports. It does not write to the ledger, and it cannot write or send through a connector either. The read-only boundary described in How Accounting Intelligence Works holds for scheduled agents too. An agent can assemble, summarize, and report. It does not post to the ledger on its own. Unattended runs are barred from a further set of actions that an interactive conversation can take. A scheduled agent cannot create a workflow, change a workflow’s status, approve or reject a record, or apply a bulk reclassification, even when the person who owns it holds those permissions. Those actions require someone present to authorize them. Connectors are held to the same boundary, and it is enforced rather than left to configuration. An unattended run receives only a connector’s read-only tools. Everything grouped under Interactive Tools on the connector, Gmail’s draft and send tools among them, is removed from the agent’s tool set before the run starts, whatever that tool’s permission is set to. A scheduled agent can search an authorized mailbox; it cannot put a message in one. Those per-tool permissions govern conversations you are present for, which is what the Interactive Tools heading means. Nothing set there widens what a schedule can do, so a scheduled connector workflow only ever takes one shape: the agent assembles the list and a person acts on it. For a worked example of the split, see how to send dunning emails with Accounting Intelligence and Gmail, where the schedule produces the list and the person sends the mail. To prove a new agent’s scope and output path before it runs unattended, see how to test a scheduled agent.
Last modified on September 14, 2026