On October 14, 2026, Anthropic updated the Claude Code desktop app: it now supports working on multiple tasks in parallel. Each task opens in a nested session, and its result returns to the main conversation. Along with this, saved environments appeared — ready-made sets of folders, SSH connections, or clouds, where each folder gets a separate Git worktree. The update works in the Claude app for macOS and Windows (Linux — in beta) on Pro, Max, Team, and Enterprise plans.

image
image

What happened

The announcement was published in the official @ClaudeDevs account on X. The app's sidebar shows which sessions are currently running, which have already finished, and which are waiting for user input. Each folder in a saved environment gets its own Git worktree — a separate working copy of the repository with its own branch, so concurrent edits do not conflict with each other. Side questions are asked with the /btw command (Cmd+; on macOS, Ctrl+; on Windows) and open in a separate tab without disrupting the main session.

Context

Until now, Claude Code worked as a single chat agent with one task and one file system, and the main risk of parallel agents is conflicting edits to the same code. A separate Git worktree per folder technically solves exactly this classic problem: each task is isolated in its own working copy and branch. Nested sessions implement the supervisor/worker pattern: the main conversation delegates subtasks to separate sessions and collects results, preserving its own state. The /btw command works with context limits, which remain a bottleneck for long agent runs. Saved environments take a step toward reproducibility: a fixed set of folders, SSH hosts, and clouds allows repeating an agent run under the same conditions, which is important not only for convenience but also for correctly evaluating agent systems.

Why this matters for the industry

Claude Code is turning from a coding agent into a full-fledged orchestrator of parallel agent sessions: nested sessions, saved environments, and worktree per folder allow one developer to handle multiple tasks in different repositories and environments — local, on SSH hosts, or in clouds — without file conflicts. Competition among coding agents is shifting from the quality of single responses to managing parallel work streams and environments. For startups, this is simultaneously a reduction in development costs due to fewer people on routine and coordination, and a blurring of the moat for thin wrappers around models. If the pattern takes hold, a wave of secondary products is likely: worktree managers, environment templates for typical stacks, and observability panels for multi-agent runs, where value shifts to summaries of session results.

Why this matters for users

If you use Claude Code, just update the desktop app (Code tab) and try running two tasks in parallel. The main practical benefit is fewer context switches between tasks and safe simultaneous code edits thanks to isolation via separate worktrees. You can save an environment with an SSH host to return to it in future sessions, and ask a side question via /btw without interrupting the main work. Availability reminder: the app is for macOS and Windows, Linux is in beta, and requires a Pro, Max, Team, or Enterprise plan.

What is still unknown / limitations

The official announcement and documentation contain no measurable results: there are no benchmarks, task success metrics, or comparison of parallel mode with manual orchestration, so confidence is justified only for the statement about the existence of features, but not for claims about productivity growth. The model did not change within the update, so the impact on its capabilities and on agent evaluation methodology is zero. The Linux version remains in beta, and its behavior may differ. Expectations that other vendors will release comparable primitives, and that the developer's working unit will become an agent session manager across multiple environments, are scenario interpretations; they need to be verified by independent measurements of parallel mode reliability and the appearance of similar features from competitors.

Sources

Author

Look at AI, editorial team