Should Your AI Agent Run 24/7? Set Operating Hours and Limits

An AI agent should run 24/7 only when the work really arrives around the clock, a late response has a real cost, and someone is accountable for what the agent does while nobody is watching. For most business roles the better default is defined operating hours, with quiet periods in which the agent queues, acknowledges and escalates instead of acting. Treat continuous operation as something each task has to earn, not a setting you switch on for the whole agent.

Key takeaways

• Availability, activity and autonomy are three separate dials. You can keep an agent reachable at all hours while limiting what it may do at night.
• Always-on pays off for urgent inbound requests and low-risk background jobs. It rarely pays off for tasks that change records, send messages or spend money.
• Unattended hours are where small faults compound: retry loops, malformed inputs and runaway usage. Limits on spend, actions and retries matter more than the schedule itself.
• A quiet period is a defined behavior (queue, acknowledge, escalate, read-only), not a shutdown.
• Review the schedule against real logs of out-of-hours activity, and change it through a normal approval path.

What does running 24/7 actually mean for an agent?

People use the phrase for three different things, and mixing them up causes most of the poor decisions.

• Availability: the agent can receive a request at any hour and respond.
• Activity: the agent runs scheduled or background work, such as checking a queue, reconciling a list or preparing a report.
• Autonomy: the agent takes actions that change something, such as updating a record, sending a message or triggering a payment, without a person approving each one.

An agent can be fully available and still have little autonomy at night. It can accept a customer request in the small hours, record it, tell the sender what happens next and hold the action until staff arrive. That is round-the-clock service with business-hours risk.

So when someone asks whether the agent should run 24/7, the useful reply is a question back: which of the three dials, for which tasks, and who is reachable if something goes wrong? If you want a broader grounding in what an agent is and does, AI Agents for Business: Leaders' Questions on Scope and Risk covers the basics of scope.

What does always-on operation actually give you?

Continuous operation has real benefits, but they attach to specific kinds of work, not to the agent as a whole. It earns its place in a few situations:

• Inbound requests where the sender expects an answer quickly, and a wait of several hours changes the outcome.
• Teams or customers spread across time zones, where the working day never fully ends.
• Monitoring tasks, such as watching a queue, a mailbox or a set of records for conditions that need attention.
• Heavy batch work that is better done when your other systems are quiet, so it does not compete with people during the day.
• Handoffs between shifts, where the agent prepares a clear summary so the next person starts informed.

Notice what these have in common. In most of them, the valuable part is noticing, recording and preparing, not deciding and acting. An agent that triages overnight and hands a ready summary to the morning team gives you most of the benefit of 24/7 operation with a fraction of the exposure.

If your team and your service provider sit in different time zones, the same logic applies to support. Working With a Remote AI Provider Across Time Zones explains how to agree on who responds when, so the agent's hours and the humans behind it line up.

What does it cost in supervision, spend and risk?

The cost of always-on operation is not mainly the extra hours of compute. It is the hours in which nobody is looking.

Supervision. During the day, a colleague notices when the agent says something odd or repeats itself. At night, nobody does. A fault that a person would catch within minutes can run for hours. Agents that work over long stretches also need clear records of what they have done and what remains, so a restart or a handoff does not lose the thread. Anthropic's engineering notes on [effective harnesses for long-running agents](https://www.anthropic.com/engineering/effective-harnesses-for-long-running-agents) describe why progress state and clean handoffs matter once an agent runs for extended periods.

Spend. Usage-based costs rise with activity, and a loop or a retry storm is activity nobody asked for. An agent that fails and retries repeatedly through the night can consume a month of budget before anyone arrives. Hard caps and alerts are the protection, and AI Agent Usage Limits and Budget Alerts: A Practical Setup walks through how to set them.

Risk. Anything that reaches the agent from outside, such as an email, a form or a chat message, is untrusted input, and it arrives at night just as it does by day. The OWASP guidance on [prompt injection prevention](https://cheatsheetseries.owasp.org/cheatsheets/LLM_Prompt_Injection_Prevention_Cheat_Sheet.html) is a useful reference for why an agent with broad permissions and no one watching is a weak position. The narrower the permissions during unattended hours, the smaller the worst case. A permission review, as described in How to Run a Permission Review for Your AI Agents, Step by Step, is the natural place to decide which actions stay available at night.

Visibility. You can only supervise what you can see. Logs, traces and alerts that show what the agent did, when and why are what turn an unattended period into a reviewable one. The [OpenTelemetry observability primer](https://opentelemetry.io/docs/concepts/observability-primer/) is a good plain explanation of the signals involved.

How do you decide? A decision guide

Work through these questions for each task the agent performs, not for the agent as a whole. The answers can differ a lot between, say, answering a status question and issuing a refund.

• Does the request lose value if it waits until the next working morning? If not, keep it in business hours and let the agent queue overnight arrivals.
• Is the action reversible without effort? If yes, it can run unattended with logging. If no, require approval or hold it for the morning.
• Does the action touch money, customer-facing messages, contracts or personal data? If yes, keep a person in the loop, or restrict the agent to preparing a draft.
• Can a person be reached within a defined time if the agent raises an alert? If no one can be reached, the agent should do less, not more.
• Do you have a spend cap, an action cap and a retry cap in place? If any is missing, do not extend the hours yet.
• Can you see, the next morning, a clear record of what happened overnight? If not, fix visibility before extending autonomy.

A task that passes every question is a candidate for continuous operation. A task that fails one of them is a candidate for a quiet period with a defined fallback. Failing several means the task belongs in supervised hours only. The NIST [AI Risk Management Framework](https://www.nist.gov/itl/ai-risk-management-framework) takes the same approach: match the level of oversight to the level of risk, and keep checking as conditions change.

What does a sensible schedule look like? An illustrative scenario

Illustrative scenario: a regional distributor uses an agent on its order desk. Customers write in from several countries, and a few messages arrive every night. The owner is nervous about both missing urgent orders and the agent sending something wrong while the office is closed.

They split the work by task and set three modes.

• Staffed hours: the agent answers questions, drafts replies, updates order notes and proposes changes. Staff approve anything that alters an order or contacts a customer about a dispute.
• Quiet hours: the agent still receives messages. It sends a short acknowledgement saying when a person will respond, tags each message by urgency, answers only from approved reference material and writes a summary for the morning. It cannot edit orders, issue credits or send anything beyond the acknowledgement.
• Escalation: if a message matches a short list of urgent conditions, such as a delivery that is blocked or a safety concern, the agent alerts the on-call person, who decides whether to act.

The agent has a daily spend cap and a limit on actions per hour, and a retry cap on every tool. If any cap is hit, the agent stops acting and raises a single alert instead of continuing. Each morning, the team reads the overnight summary, and at the end of the first few weeks the owner reviews what the quiet-hours mode caught and missed before changing anything.

The agent is reachable all night, yet the exposure is the same as a well-briefed receptionist who can take messages but not make commitments.

How do you set operating hours, quiet periods and limits, step by step?

Use this plan when you set up an agent, or when you revisit an existing one.

• Step one: list the tasks. Write down every action the agent can take, tool by tool, and mark each as read, draft or change.
• Step two: classify by urgency and reversibility. Note which tasks lose value when delayed and which cannot be undone.
• Step three: define the windows. Decide the staffed hours, the quiet hours and any days when the agent runs in a reduced mode, such as holidays.
• Step four: write the quiet-period behavior. For each window, specify what the agent may do (queue, acknowledge, summarize, answer from approved material), what it must not do and what triggers an escalation.
• Step five: set the limits. Define a spend cap, an action cap per period, a retry cap per tool and a maximum run time for any single job. Decide what the agent does when a limit is hit: stop, report once and wait.
• Step six: name the people. Assign an on-call contact for escalations, an owner for the schedule and a deputy for both. Agree on the expected response time.
• Step seven: add a stop control. Make sure an authorized person can pause the agent or restrict it to read-only quickly, without needing the builder. The guidance on stopping an AI pilot without wasting the work explains pausing safely.
• Step eight: make the night visible. Confirm that logs, alerts and a morning summary exist, and that someone reads them.
• Step nine: trial the schedule. Start with narrow overnight autonomy and widen it only when the logs show the limits were respected and the escalations were useful.
• Step ten: schedule a review. Compare what the agent did out of hours against what you intended, and adjust the windows or limits.

Who may change the hours and limits is itself a decision. Which AI Agent Settings Should Business Owners Control? is a useful reference for separating the settings an owner should hold from those best left to the operator. The OWASP [authorization guidance](https://cheatsheetseries.owasp.org/cheatsheets/Authorization_Cheat_Sheet.html) supports the same principle: grant the least access needed, and check it at the time of each action.

How does Be Human handle operating hours?

Be Human treats operating hours as part of the agent's role definition, set during the first engagement alongside its memory, tools and approval boundaries. Each role gets a defined schedule, quiet-period behavior, limits on spend and actions, and a named escalation path, all under the client's control. Managed operations then include monitoring of out-of-hours activity, recurring evaluations of what the agent did, permission review and controlled changes to the schedule when the work changes.

Frequently asked questions

Is it safe to let an AI agent run overnight without anyone watching? It can be, if the agent's overnight permissions are narrow, its spend and action limits are firm, and its activity is logged and reviewed each morning. Reading, summarizing, acknowledging and escalating are generally safe to run unattended. Actions that are hard to undo should wait for a person.

Does running an agent around the clock cost more? Often, yes, since usage-based costs follow activity, and quiet hours with fewer actions tend to cost less. The larger cost is usually unplanned activity, such as loops and retries, which spend caps and alerts exist to contain. Compare the cost of extended hours with the value of faster responses for the specific tasks that need them.

What should an agent do during a quiet period? It should keep receiving requests, acknowledge them, sort them by urgency, answer only from approved material and prepare a summary for the team. It should not change records, spend money or send substantive messages. A short list of urgent conditions should trigger an alert to a named person.

How often should we revisit the operating hours? Review them after the first few weeks of live operation, then on a regular cycle and whenever the work, the team or the agent's tools change. Use the overnight logs as evidence: what arrived, what the agent handled, what it escalated and what it missed. Changes should go through your normal approval path, as described in How to Request and Approve Changes to a Managed AI Agent.

Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.