Idle Auto-Pause

Idle Auto-Pause

Idle auto-pause is an automated policy that suspends a running agent after it has gone a defined period without receiving or processing a task. Rather than requiring an operator to manually stop every agent that finishes its work, the platform detects inactivity and transitions the agent to a paused state on its own, freeing the underlying compute for other use.

How it works

An idle timer tracks the time since an agent last received a task or produced output. Once that timer crosses a configured threshold, the platform triggers a pause transition, similar to a manual pause but initiated by the system rather than a person. The agent's container and any in-progress state are stopped or suspended, and its data is retained so the agent can be brought back to activity later without loss of context.

Why it matters

A fleet of long-lived agents behaves differently from a batch of short-lived jobs: agents are often provisioned once and kept around indefinitely so they retain conversation history, credentials, and working state. Left running continuously, an idle agent still occupies a container, memory, and often a reserved slot against provider rate limits, even though it is doing no useful work. Idle auto-pause closes that gap: it lets an operator provision many agents without paying the ongoing compute cost of every agent that happens to be between tasks.

Auto-pause vs archival

Idle auto-pause is a short-term, automatic response to inactivity, while archival is typically a longer-term, often manual decision to set an agent aside. A paused agent is expected to resume soon and wakes automatically on the next task; an archived agent is expected to remain inactive until an operator explicitly restores it.

Idle auto-pause in Agenhood

In Agenhood, idle agents pause automatically to save compute, and they wake automatically the next time a task is submitted to them. An operator does not need to build separate tooling to detect and shut down inactive agents, and does not need to remember to restart an agent before assigning it new work; submitting a task to a paused agent is enough to bring it back to a running state.