Anthropic published a guide to always-on agents on November 13, 2026. It describes how software can notice new work, process it in short runs and keep enough information to pick up later, even when no chat window is open.
An event starts each run
Anthropic describes an agent that sleeps between runs. A message, scheduled task or background event can wake it. The guide recommends an event inbox so the application can record that work has arrived before giving it to the worker responsible for the next run.
An inbox also makes duplicate deliveries visible. A reply to earlier work may resume the same session, while an unrelated event can start a fresh one. Timers are held separately until they come due, so the inbox is reserved for work that can be picked up now.
What carries between sessions
The guide recommends keeping durable memory separate from the incoming events. A worker can start a session with user instructions and the notes it needs rather than load every earlier exchange. Anthropic suggests storing a pointer when a fact can be looked up reliably somewhere else, such as the current state of a ticket in an issue tracker.
Approvals and messages still need design
Permission rules matter because the agent may be active when the user is elsewhere. Anthropic says the application must decide when a tool action can proceed and when it must wait for approval. The outbox applies the same separation to delivery: the agent can decide what to report without controlling exactly when the user receives it. The guide also recommends testing when messages are helpful so that a system notices problems without turning every routine run into an interruption.
