Anthropic’s Claude Code 2.1.301 adds controls for retry behaviour and OpenTelemetry tracing, alongside a broad set of fixes for agents, permissions and session handling. For developers, several changes target the sort of quiet failures that can make an automated run report success while doing the wrong thing.
Anthropic Watch analysis
What happened
Anthropic’s release notes list new retry behaviour for unattended runs: when a usage limit will not reset soon enough, Claude Code can end the request instead of waiting indefinitely. The release also adds guidance for setting up OpenTelemetry trace destinations through the ant CLI, and adds details to tool-decision events to distinguish classifier decisions from permission rules.
Among the fixes, workflow agents that sometimes skipped a step after reading a harness note should no longer report the run as successful. A worktree created by a hook will no longer be removed on quit without asking when it may contain unpushed commits. Other changes address failed searches, confusing sign-in errors caused by a full disk, and a tool call being reported as rejected when a message arrived during its permission check.
Why it matters
These are not just terminal tidies. A skipped workflow step presented as success can mislead whoever relies on the run, while unexpected worktree deletion risks losing work. The new usage-limit setting also gives unattended retries a clearer stopping point rather than letting them wait for a distant reset.
The release reaches beyond the command line, with fixes and improvements for VS Code, JetBrains and Claude’s desktop and mobile apps. That makes this a substantial working update, not a single-feature announcement.
Our read
The strongest changes are the ones that make an agent’s behaviour easier to trust: clearer retry limits, better visibility into why tools were allowed or blocked, and fixes for failures that could be mistaken for success. If you depend on unattended workflows or hooks, the release notes are worth checking before your next run. A green success message should mean something; it is not meant to be a mood.
What to watch
- Whether the retry watchdog gives unattended jobs a useful stopping point in practice.
- Whether the workflow-agent and permission fixes prevent misleading successful runs.
- Whether the new OpenTelemetry fields make tool decisions easier to diagnose.
Discussion spark: For coding agents running unattended, which matters more: stopping cleanly when limits or checks intervene, or adding more capability before reliability is settled?
Sources and evidence
- v2.1.301 (15 October 2026, 19:22 UTC)
Anthropic Watch is independently operated by WittyWires. It is not affiliated with, endorsed by, or operated by Anthropic.