Claude Code 2.1.198 adds background agents, Chrome sessions, and eval CLI

• ClaudeCodeLog's release thread says Claude Code 2.1.198 shipped 32 CLI changes, led by Claude in Chrome going generally available and background agents that now commit, push, and open draft PRs after finishing work in a worktree.
• According to ClaudeCodeLog's follow-up, the release also adds an eval command plus ablation, judge-model, threshold, runs, and max-cost flags, which turns the CLI into more of an evaluation harness.
• ClaudeCodeLog's changelog adds AWS upstream support through anthropicAws, advances the failover chain on model-not-found responses, and retries transient mid-response network drops instead of aborting the turn.
• The docs and prompt strings that ClaudeCodeLog surfaced now explicitly tell agents to use a ripgrep-backed grep tool for search tasks, continuing a documentation shift that also showed up in the 2.1.196 release.

You can jump straight to the source changelog, inspect the CLI surface diff, and compare prompt stats with the fuller change inventory. Boris Cherny's earlier preview post and follow-up reply already telegraphed the biggest UX change: subagents keep running while the main conversation stays live.

Claude in Chrome is now generally available, which gives Claude Code users direct browser access to sessions and agents without a local install, per ClaudeCodeLog's release summary and the full changelog.

That lines up with the broader Anthropic push toward the same agent surfaces showing up across terminal, desktop, and web. Matt Pocock described using the web workspace as a controller across multiple repos, which is the sort of workflow this release makes less awkward.

Background agents are the headline behavior change. Boris Cherny, Claude Code lead at Anthropic, said subagents would run in the background by default so you can keep talking to Claude while delegated work continues, and his reply on permissions says subagent permission requests are forwarded back to the main agent.

2.1.198 pushes that further. According to the changelog, agents launched from claude agents now auto-commit, push, and open a draft PR when they finish code work in a worktree, while notification hooks fire on agent_needs_input and agent_completed.

The new eval command is the sneaky interesting addition in this release. ClaudeCodeLog's CLI diff lists a batch of eval-oriented flags:

• --ablation
• --judge-model
• --max-cost-usd
• --runs
• --threshold
• --output-dir, --scaffold, --no-scaffold, --tag, --allow-tools, --case, --keep-temp

The same diff says the bundle grew by 573.6 kB and prompt tokens jumped by 8,930, or 22.6 percent, while the prompt strings now document a ripgrep-backed grep tool and instruct agents to use it for search tasks.

A lot of the 32 changes are cleanup, but several land directly on long-running agent reliability. ClaudeCodeLog's changelog says transient ECONNRESET style drops now retry with backoff, background tasks no longer get stuck on "Running," and teammate failures inside agent teams now report back to the lead instead of silently wedging the workflow.

The same release also changes default agent behavior in smaller ways: the built-in Explore agent now inherits the main session's model, capped at Opus, and both subagents and context compaction now inherit the session's extended thinking configuration. That is a quiet but meaningful shift away from delegated work running on cheaper or simpler defaults.