Anthropic’s Claude Code has rolled out a new Monitor tool designed to let AI agents listen for streamed events in the background instead of repeatedly polling for updates inside an agent loop. Noah Zweben, product manager for Claude Code, said on X that the feature is now live and allows Claude to run scripts in the background while continuously watching standard output. Each line of output is turned into an event notification in real time, which then wakes the agent to handle the next step.
Event-driven wakeups replace repeated status checks
Before this release, Claude often had to keep sending requests to check whether an external program had finished running. That polling model consumed tokens while adding stretches of idle waiting. Monitor changes that flow. A background script pushes events outward, and Claude only wakes when a notification arrives, removing the need for constant status checks.
Zweben described the feature as an important tool for saving tokens. In practical terms, it shifts part of the agent workflow away from repeated model queries and toward an event-driven setup built around streamed output.
Designed for logs, PR tracking, and long-running listeners
Claude Code said Monitor already supports several use cases, including watching application logs for errors, using scripts to track Pull Request status, and handling other background jobs that need continuous monitoring. These tasks tend to run for extended periods and do not always change state on a fixed schedule, which makes polling expensive from a token perspective.
The company also drew a line between Monitor and the Bash tool’s run_in_background parameter. The Bash option is better suited to one-off jobs that simply run in the background until completion. Monitor is intended for tasks that need continuous streamed events and ongoing agent reactions as those events arrive.
Standard output becomes a live event stream
The key mechanism is the conversion of standard output into a stream of event notifications, line by line. Instead of waiting until a script finishes and then reading everything at once, the agent receives updates as execution unfolds. Claude Code did not provide additional performance figures in the announcement, but the stated goal is clear: cut unnecessary polling and reduce token consumption tied to waiting loops.