Claude Code 2.1.198: browser agent GA, and the handoff moves into the draft PR

The June changelog wave I covered in Claude Code hardens orchestration and in the post on MCP retries and the agent panel was about reliability: limits, timeouts, retries. Version 2.1.198 of 1 July is different. Two entries change not how stably the agent runs, but where your work with it happens. Claude in Chrome is now generally available, and background agents commit, push, and open a draft PR instead of stopping at the end and waiting for you.

What is in version 2.1.198?

Two entries shift the workflow, and a third delivers the signal path to go with them. First: Claude in Chrome, the browser agent, is now generally available. The changelog gives it exactly one sentence. Second: background agents launched from claude agents now commit, push, and open a draft PR when they finish code work in a worktree, instead of stopping to ask. Third: sessions that need input or finish now fire the Notification hook, with agent_needs_input and agent_completed as the events.

What does general availability of Claude in Chrome mean?

It means a change of status: the browser agent is no longer a preview but a regular part of the product. A single sentence in the changelog, but it changes what Anthropic expects of the tool. A preview is allowed to fail; a generally available feature is meant to carry weight. In practice this makes the browser a surface where an agent acts, not just reads. My assessment stays the same as with every new action space. Reach grows faster than verifiability, because what an agent does in a browser leaves fewer traceable marks than a diff in a repository. Which is exactly why I find the second entry in the same release more interesting: it shows the counter-model, work that ends as a reviewable artifact.

Why is the draft PR the better handoff?

Because it moves the review to where it happens anyway. Until now, a background agent finished its work, stopped, and asked what to do next. That sounds polite, but it is expensive: the run sits idle until you notice, and the answer lands in a chat nobody but you can see. Now the agent delivers an artifact. Commit, push, draft PR. The word draft carries the architecture: nothing gets merged, nobody has approved anything. The agent produces a proposal in the format your team already reviews, with a diff, a history, and a comment thread. Your decision does not disappear; it moves from the chat window into the review. And the run no longer spends hours waiting just because nobody happens to be watching.

And when the agent does need you?

Then it speaks up instead of standing still. The new Notificationhook fires when a session needs input or finishes. What happens next is up to you: a desktop notification, a message to the team, a line in a log. The point is that you no longer have to watch a panel to know when you are needed. Another entry in the same release is worth noting: subagents now treat messages from the agent that launched them as normal task direction, but an agent’s message is still never treated as the user’s approval. Agents may direct agents. Only the human approves. That is exactly the boundary you have to pin down before letting agents commit and push unattended.

What does this mean day to day?

The workflow shifts: launch, get notified, review the draft PR. That makes the review the load-bearing checkpoint, and it has to carry that weight. A draft PR you wave through unread is not a review, it is a signature under someone else’s work. It is the same line as in agentic engineering: the agent’s capability grows, the verification stays your job. Version 2.1.198 does not take that job off your hands. It only moves the place where you do it, from a mid-run stop in the chat into the pull request. That is a better place. But only if you actually look.

Sources

• Claude Code: Changelog