Always-on Agent
An always-on agent is a computer program with artificial intelligence that runs permanently in the background and decides for itself when to become active. Unlike a chatbot, it does not wait for a question, but instead observes events and acts on its own initiative.
Most programs with artificial intelligence work according to the question-and-answer pattern. You type something in, the program computes, it answers, and then the process ends. An always-on agent reverses this pattern. It runs around the clock on a server and continuously observes what is happening: new emails, changed prices, an error in a system. When a condition occurs that it considers important, it becomes active on its own. The English expression “always on” means exactly that: permanently switched on, never dependent on an external trigger.
From tool to employee on call
The difference sounds technical, but it is economically significant. A chatbot is a tool: it is only useful as long as someone sits in front of it and operates it. A permanently running agent, on the other hand, takes over a task completely, even at night and on weekends. Companies therefore no longer calculate in terms of answers, but in terms of completed processes.
This is precisely why investors have been following the topic so closely since 2024. A program that only answers saves an employee some time. A program that fully takes over a recurring task replaces work hours. Providers often charge fixed monthly prices per agent for this instead of cent amounts per request.
But there is a downside. An agent that acts without asking can also make mistakes without asking. If it runs permanently, it may repeat this mistake a thousand times over before anyone notices. That is why such systems are usually given hard limits: certain actions may only be approved by a human.
Triggers, memory and tools
Technically, an always-on agent consists of three parts. The first is the trigger. This can be a schedule, for example every minute, or an event, for example an incoming message. The trigger wakes the agent up; otherwise it sleeps and costs almost nothing.
The second part is memory. A normal chatbot forgets everything after the conversation. An agent that works over weeks must remember what it has already done. To do this, it writes its results to a database and reads them back in when it next wakes up. Without this memory, it would constantly start the same work over again.
The third part is the tools. A language model can, on its own, only generate text. For it to have any effect, it is given access to programs: a search engine, a calendar, an inventory management system. The model then decides which tool to call and reads the result back in. You can think of it like an intern who can think, but only becomes capable of acting through passwords and access rights.
Where such agents are already at work
In everyday life, one usually encounters them unnoticed. One agent monitors the servers of an online shop and initiates a restart before customers notice anything. Another reads a customer service inbox, sorts inquiries, and answers simple ones itself. In software development, agents check new program code and suggest corrections without anyone asking them to.
In the financial industry, similar systems are running that continuously monitor news and prices and send an alert when unusual patterns appear. It is important to distinguish this from automated stock trading. The agent prepares and reports; the decision is usually made by a human.
In news reports, the term often appears together with “agentic AI.” A common misconception here is to consider every agent an always-on agent. Many agents only work through a single task and then terminate themselves. It is only permanent operation with its own trigger that makes the difference.