How AI keeps working when nobody is typing

An AI system can continue a job without someone typing another message. Software can supply its next input, execute a requested action and return the result for another decision. Add a schedule and a saved record of the work, and the same arrangement can support an ongoing responsibility: checking information, responding to changes and producing an outcome when it is needed.[3][4][5]

This is the practical meaning of autonomy: how much a system can decide and do within its assigned role. Consciousness asks whether there is subjective experience. Automatic initiation and purposeful-looking behaviour do not, by themselves, answer that question. A paper by Patrick Butlin and colleagues instead proposes assessing indicators derived from scientific theories of consciousness, while recognising substantial uncertainty about those theories.[1][2]

The next prompt can arrive automatically

Consider an illustrative travel assistant assigned to check a commuter’s morning journey. At a specified time, a scheduler starts the job. The assistant reads the relevant service information, checks whether the planned journey is affected and prepares alternatives if necessary. Nobody needs to wake it with a new message each morning. The user has already supplied the standing instruction; the application supplies the occasion and current information.[4][3]

LangChain’s scheduling documentation describes this arrangement concretely: a server can create an assistant thread on a timetable and send it a specified input. The instruction can stay the same on every run while the information retrieved during the run changes. That is an important source of apparently spontaneous behaviour. A fresh observation can come from an ordinary scheduled process.[4]

Choosing a route through the work

There are different degrees of freedom inside that process. A fixed workflow might always fetch a status report, summarise it and send the summary. An agent can choose additional steps in response to what it finds: inspect a second source, compare alternatives or ask for missing information. Anthropic’s engineering guidance distinguishes predetermined workflow paths from systems in which the model directs the next operation.[3]

The distinction is useful because automation can be valuable without much autonomy. A predictable task may need a dependable sequence. A changing task may benefit from a model choosing the sequence. Calling both “AI agents” can obscure the actual capability being purchased: following a procedure, selecting a procedure or constructing one as the work unfolds.[1]

A continuing job needs a continuing record

A long-lived responsibility also needs information to survive between executions. LangGraph’s current documentation separates a checkpoint of the present thread from a store of information used across threads. In the travel example, one record might preserve an interrupted search; another might retain the normal destination and accessibility preferences. A new run can use those records without keeping every earlier conversation in its active context.[5]

Anthropic demonstrated a complementary technique for extended coding work in November 2025. Its system left a progress file, a feature list and version-control history for later sessions. A new session could inspect what had been completed and continue with an unfinished feature. The practical advance was a workable handover between periods of model execution. Merely repeating the original request had led to unfinished changes and premature declarations of completion.[6]

Saving such a record does not necessarily train the underlying model. The OECD distinguishes autonomy from adaptiveness after deployment: acting with less intervention and changing through learning are separate dimensions. An assistant can become more useful because it has better information about a job, even when its model weights stay unchanged.[1][5]

Independence can be observed

Anthropic’s February 2026 study illustrates how this capability can be measured in an actual product. In Claude Code, the median turn lasted about 45 seconds, while the 99.9th-percentile turn duration rose from under 25 minutes to over 45 minutes between October 2025 and January 2026. These were measured periods of work before the system stopped, asked a question or was interrupted—not a claim that ordinary users routinely delegated 45-minute jobs.[7]

The study also makes clear why duration alone is incomplete: a better system may finish the same job sooner. What matters to the person delegating work is whether the necessary decisions and actions happen successfully. Recurring autonomy adds another dimension: the application must recognise the next occasion, recover the relevant state and start again. That is the machinery that turns a useful conversation into an ongoing service.[7][4][5]

Report an error or suggest a correction