v2.1.198

Source: Anthropic Claude Code Date: 2026-07-01 URL: https://github.com/anthropics/claude-code/releases/tag/v2.1.198

Summary

Claude Code v2.1.198 makes background subagents the default execution mode and, on finishing code work, has them automatically commit, push, and open a draft PR “instead of stopping to ask.” The Explore agent now inherits the parent session’s model (capped at Opus) rather than running on Haiku, Claude in Chrome goes GA, and the /agents wizard is removed in favor of direct editing of .claude/agents/.

Implications

This is the opening move in what the landscape thread logs as the “autonomy sprint” — the first of a four-release run (v2.1.198→201) that strips asks out of the machine-to-machine loop while holding one boundary explicit: an agent’s message is never treated as the user’s approval. Feeds the un-delegatable-signature thread (permission doesn’t delegate, and per the following days’ releases, neither does verification) and the agentic engineering / harness-reliability thread this radar tracks across CLI releases (Codex, Gemini CLI, Vibe, Beads) as autonomy features roll out ahead of the reliability work needed to trust them unattended.