--- title: What does an "always-on" AI agent actually do, and what should it cost? author: Giacomo Balli published: 2026-11-13 url: https://BigBalli.com/blog/always-on-ai-agents-what-they-really-do tags: AI Agents, Automation, Monitoring, AI Costs, Vendor Evaluation --- # What does an "always-on" AI agent actually do, and what should it cost? *By Giacomo Balli* | *Published: November 13, 2026* **Short answer:** An always-on AI agent is usually a cheap program that listens all the time and wakes the AI model only when an event needs judgment. A model pays for every word it reads, so one that reads everything bills you for quiet days. Filter first, run the model on hits, and let it act alone only on a written list of reversible actions. An "always-on" AI agent is usually a cheap program that stays awake and listens for a new email, a webhook, a log line or a schedule, and wakes the AI model only when something needs judgment. That design matters for cost. A model should cost money while it works, never while it waits. If a vendor's agent bills you for watching a quiet system, it is built wrong. 1. **Listener, always on**Reads every new line, message or event. Plain code, no AI, costs nearly nothing. 2. **Filter**Fixed rules and thresholds decide what is worth a look. Most events stop here. 3. **Model, on demand**Gets the flagged event plus its context and judges it: real problem or noise. 4. **Act or ask**Fixes what it is allowed to fix alone, and asks the owner about everything else. The listener never sleeps. The model wakes only on a hit. ## What does "always on" mean when a vendor says it? When a vendor says "always on", it means the agent answers whenever something happens. A model thinking around the clock is rare. Slack assistants, email agents and incident-response bots all work the same way: an event arrives, the model handles that one event, and it goes idle until the next one. OpenAI's Agents API, which offers ready-made templates for an incident-response agent and a Slack teammate, keeps durable sessions that can stay open for days. The model inside them runs, and bills, only while there is an event to process. Usually something pushes the event to the agent: a new message, a payment webhook, a file landing in a folder or a rule firing in a log wakes the agent. Polling every few seconds works too, but each poll that finds nothing should cost close to zero. If it runs the model, it does not. ## Why not let the model read everything as it arrives? Because a model pays for every word it reads, whether the word matters or not. Take a real example. On November 12, the web server behind this site wrote 207,612 lines to its main access log, about 60 MB of text. Fed to a language model line by line, that is on the order of 15 million tokens a day. Of those lines, 1,604 were obvious attack probes: bots looking for leaked password files, WordPress logins and exposed code folders. That is under 1 percent. The other 99 percent were ordinary page views the model would read only to conclude that nothing happened. A model that stays running costs more. To "remember" what it has seen, a running session keeps the history in its context, and every new turn sends that history again. Cost per line climbs the longer it runs. The tools that trim old context to keep costs down also throw away the memory that was the reason to keep the session open. Running the model more often does not fix this either. A daily security review that reads the whole day's log costs the same tokens as one that reads it in 1,440 one-minute slices. Frequency changes how fast you hear about a problem. The total text the model reads stays the same. The only lever on cost is how much text reaches the model. ## How should an agent that watches your business be built? Put a dumb, always-on filter in front and wake the model only when the filter fires. For the server log above, that looks like this: - **The listener** follows the log files as they grow and remembers where it stopped. Free tools such as fail2ban and CrowdSec already do this for security and block repeat offenders on their own. - **The filter** holds known-bad patterns (probes for password files, failed logins, bursts of server errors) and rate rules, such as one address making hundreds of requests a minute or a new client hitting an admin page. - **The model** receives the flagged event plus the 50 lines around it and answers the questions rules cannot: is this a real attack or a scanner, did any request succeed, what else did that address touch? - **The action** comes from a fixed list. Blocking an address or restarting a service can happen alone. Anything that sends, deletes, pays or touches a customer waits for a yes from the owner, by email or a one-tap approval link. On a quiet day the model reads nothing. On a noisy day it reads a few dozen events instead of 200,000 lines, and it reacts in seconds instead of at the next daily report. A daily run still earns its place, reading the filter's summary to catch slow patterns no single rule sees. Two details separate a safe design from a risky one. First, the text an agent watches is often written by strangers: a log records the address and browser name an attacker chose, and an inbox holds whatever anyone sends. Text like that can carry instructions aimed at the model, so the step that reads it should only be able to recommend. Second, the listener needs a heartbeat. A watcher that died at 3 a.m. looks exactly like a quiet night. ## When does a model need to run continuously? Only when nearly every new piece of input needs judgment, quickly, in light of everything before it. Three cases qualify: - **Live conversation.** A voice agent on a phone line, or an assistant that follows a sales call and suggests the next question, has to weigh every turn the other person takes. Each turn needs judgment and the earlier turns are the context. - **A long task in progress.** An agent spending three hours on a data migration or a software build is always on because it is busy working. What it has learned along the way is the value. - **Continuous control.** Robotics, games and live trading act on a stream with no quiet stretches. These rarely use language models, because a language model is too slow and too expensive per decision. Logs, metrics, inboxes, prices, reviews and stock levels fail the test. Most of what arrives needs no judgment at all, so a filter should see it first. ## What should you ask a vendor selling an always-on agent? QuestionGood answerWarning sign What wakes the model?Specific events or rules you can read and change"It monitors everything in real time" What does a quiet day cost?Close to nothingA flat usage bill that does not move with activity What can it do without asking?A short written list of reversible actions"It handles it" with no list How does it ask you?A message with the evidence and a one-tap approve or declineA dashboard you have to remember to check How do you know it is alive?A heartbeat that alerts when it stopsSilence, which looks the same as health ## Related guides - [Who is liable when a business's AI agent causes harm?](https://BigBalli.com/blog/who-is-liable-when-your-ai-agent-causes-harm) - [Does my business need a full AI chatbot model or a small decision model?](https://BigBalli.com/blog/do-you-need-an-llm-or-a-small-decision-model) - [When is an AI API wrapper enough, and when do you need more?](https://BigBalli.com/blog/build-vs-buy-ai-api-wrapper) - [OpenAI: choosing an agent runtime](https://developers.openai.com/api/docs/guides/agents) - [fail2ban: open-source log watcher that blocks repeat attackers](https://github.com/fail2ban/fail2ban) ## Key takeaways - "Always on" describes the listener. The model should run only when an event needs judgment. - A model pays for every word it reads. One small business web server produced 207,612 log lines in a day, and under 1 percent mattered. - Running the model more often changes response time and leaves cost unchanged. Cutting what reaches the model is the only cost lever. - Give the agent a written list of what it may do alone, and route everything else to the owner for a yes. - Continuous models make sense for live conversation, long tasks in progress and continuous control, and almost never for watching. ## Frequently asked questions ### Is an always-on AI agent expensive to run? It should not be. A well-built agent costs almost nothing while nothing happens, because a plain program does the watching and the model runs only on flagged events. If the bill stays flat whether your systems are quiet or busy, the model is reading everything, and you are paying for it. ### Can an AI agent monitor my systems and fix problems on its own? Yes, within limits you set in writing. Reversible actions such as restarting a service, blocking an address or retrying a failed job can run alone. Anything that sends a message, deletes data, spends money or reaches a customer should wait for your approval. ### Should a monitoring agent run every minute or once a day? Neither, if the model reads raw data each time. Both read the same amount of text in total. Let a cheap filter watch continuously and wake the model only when it flags something, then keep a daily run for the summary. ## About the author Giacomo Balli is an independent technology advisor in San Francisco. He has built software and mobile apps since 2010 and reviews software, vendor, and contract decisions for owners before they commit the money. ## Disclosure Giacomo Balli sells fixed-fee independent reviews of technology decisions, including AI tools and vendors. He does not build software for clients, resell platforms, or take referral fees. No company named on this page paid to be mentioned. This page is general guidance and not legal advice. ## Questions this answers **Is an always-on AI agent expensive to run?** It should not be. A well-built agent costs almost nothing while nothing happens, because a plain program does the watching and the model runs only on flagged events. If the bill stays flat whether your systems are quiet or busy, the model is reading everything, and you are paying for it. **Can an AI agent monitor my systems and fix problems on its own?** Yes, within limits you set in writing. Reversible actions such as restarting a service, blocking an address or retrying a failed job can run alone. Anything that sends a message, deletes data, spends money or reaches a customer should wait for your approval. **Should a monitoring agent run every minute or once a day?** Neither, if the model reads raw data each time. Both read the same amount of text in total. Let a cheap filter watch continuously and wake the model only when it flags something, then keep a daily run for the summary. --- **Tags:** AI Agents, Automation, Monitoring, AI Costs, Vendor Evaluation --- *View the [HTML version](https://BigBalli.com/blog/always-on-ai-agents-what-they-really-do) of this post.*