How is an always-on agent different from a chatbot or a scheduled job?

A chatbot only works while a person is sending it messages, and a scheduled job runs the same predefined steps every time its timer fires. An always-on agent runs with nobody in the conversation, like a job, but its model chooses the steps on each run, working toward a standing goal and acting on new events as they arrive.

The chatbot difference is about who starts the work. Harrison Chase’s January 2025 LangChain post “Introducing ambient agents” notes that ChatGPT and other chatbots wait for you to start the conversation, and you can only run one conversation at a time. His “ambient agents,” the same shape under another name, listen for events such as incoming email and can act on several at once.

The scheduled-job difference is about who writes the steps. Anthropic’s December 2024 post “Building effective agents” calls an LLM system whose model calls and tools run through predefined code paths a workflow, so by that definition a cron job that calls a model is still a workflow. Choosing its own path is what separates an agent from a program that calls a tool, and it means two runs from the same trigger can do different things.

Products blur the line. As of October 2026, OpenAI’s ChatGPT documentation says scheduled tasks run in the background on a timer or, on eligible plans, on a supported Gmail, Slack, or GitHub event. So the test that holds is whether each run follows steps written in advance or decides them as it goes.