Claude Code v2.1.198: Background Agents Now Open PRs Automatically, /dataviz Skill Arrives

Anthropic released Claude Code v2.1.198 on July 1, 2026 — a release that delivers a series of changes aimed at more autonomous development agents and better infrastructure reliability. The update covers more than 30 bug fixes alongside several new features that change how background agents behave.

How Do Background Agents Now Close the Development Cycle on Their Own?

The most important change in this release concerns the behavior of agents launched with the claude agents command. Previously, background agents working in an isolated worktree would stop and wait for the user to confirm when they finished. From v2.1.198 onward that changes — agents now automatically commit changes, push the branch, and open a draft pull request without interrupting the workflow.

Two new events that fire the Notification hook have also been introduced: agent_needs_input and agent_completed. This allows external integrations and notifications to react at the exact moment an agent requires human input or a job finishes — without any manual status tracking.

The /agents wizard that previously walked users through agent setup has also been removed. Subagents are now configured conversationally or through files in the .claude/agents/ directory, which unifies the interface and reduces visual complexity.

Claude in Chrome Generally Available

Claude in Chrome, which had been in beta or preview status, reaches general availability (GA) with this release. Users can now use Claude directly inside the Chrome browser without special requests or waiting, extending the assistant’s reach beyond the terminal and IDE environment.

New /dataviz Skill

The release introduces a built-in /dataviz skill designed for chart and dashboard design. The skill ships with a runnable color-palette validator that checks visual consistency and the accessibility of chosen colors — useful for development teams generating analytical views or reports.

Gateway: AWS as a New Upstream Provider

At the infrastructure level, Claude Platform on AWS (anthropicAws) has been added as an upstream provider in the gateway layer. Additionally, situations where a model was not found (model-not-found) no longer terminate with an error — instead, the system automatically advances to the failover chain and tries an alternative provider. Both improvements increase operational resilience for users running in production environments with multiple providers.

Agents Inherit Extended-Thinking Configuration

Subagents and context compaction processes now inherit the extended-thinking configuration of the session that launched them. The same applies to the built-in Explore agent, which previously used the Haiku model — from this version it inherits the main session’s model, capped at Opus. This change improves consistency and output quality for complex delegated tasks.

Reliability and Bug Fixes

The release also brings a series of reliability improvements that had been a source of complaints. Transient network errors of the ECONNRESET type are now automatically retried with exponential backoff instead of failing immediately. Background tasks no longer get stuck in a “Running” status. In agent teams, failed agents now report their status to the orchestrator agent, and sending a message to a stuck agent triggers an immediate retry.

On the interface side, the /diff panel now refreshes when switching branches or committing, and Markdown tables render correctly in fullscreen mode. Syntax rendering has also been improved through an upgrade to highlight.js 11.

Users in AWS/Mantle sessions will no longer hit a dead end with the message “Please run /login,” and the /desktop command now works correctly after entering and exiting a worktree.

Claude Code v2.1.198 is available immediately via the standard update mechanism.