Claude Code v2.1.98 shipped the Monitor tool, and it changes how the agent handles long-running processes. Instead of polling on an interval to check if something happened, Monitor launches a shell command in the background and streams each line of stdout back to Claude as an event. If nothing prints, Claude does nothing and burns zero tokens. The moment something relevant appears, Claude wakes up and reacts.
The practical difference is significant. You can tell Claude to run kubectl logs -f | grep --line-buffered error and it will sit quietly until a matching log line appears, then start working on a fix. Same idea works for dev servers, CI pipelines, deploy scripts, or any process that writes to stdout. Lines arriving within 200ms of each other get batched into a single notification, so multi-line stack traces group naturally.
This is event-driven agentic development. Before Monitor, watching a background process meant either running it in the foreground (blocking Claude) or looping with sleep intervals (wasting tokens on empty checks). Monitor eliminates both problems. Your dev server, log tail, or build process runs in the background while Claude continues other work, only interrupted when something actually needs attention.