Claude Code 2.0: Massive Upgrade with Agent Loops

Quick Overview

Claude Code 2.0 introduces massive upgrades, centering on the new /loop command for session-scoped scheduled tasks within the terminal, which allows for persistent, self-improving automation tasks that survive restarts, contrasting with Cowork's cloud-based scheduling that runs in a VM sandbox.

Key Points: The new /loop command enables scheduling recurring tasks directly within the Claude Code terminal session, supporting intervals up to 3 days. Desktop scheduled tasks are persistent, surviving restarts, and execute on the user's local machine, accessing local files, MCP servers, skills, git, and the terminal. Cowork scheduled tasks run in a VM sandbox on the cloud infrastructure, best suited for non-developer tasks like email briefings, file cleanup, and using web search/plugins. The CLI /loop command is session-scoped, meaning tasks are killed if the terminal session closes, the computer sleeps, or after a 3-day auto-expiry. Desktop scheduled tasks follow a self-improving loop pattern: read state.json, do work, write updated state.json, allowing them to remember past runs and fix errors over time. A key permission reality check exists across all platforms: desktop tasks stall on permission prompts, CLI needs permissions, and fixes often require manually allowing tools first. The scheduling surface is mapped across two axes: Task Type (Business/Life vs. Code/DevOps) and Time Horizon (Short-lived/This session vs. Long-running/Indefinite).

Context: The video introduces significant new features for Claude Code 2.0, focusing heavily on the introduction of scheduled tasks via the command in the terminal, contrasted with the existing cloud-based scheduling available in the Cowork tab. The core difference highlighted is execution environment: local machine for Desktop/CLI tasks versus a VM sandbox for Cowork tasks, leading to different access capabilities and use cases for each.

Detailed Analysis

The video details the new scheduling features in Claude Code 2.0, primarily focusing on the CLI command for session-scoped scheduled tasks within the terminal, and contrasting this with the existing Cowork scheduled tasks. Desktop scheduled tasks, initiated via in the Desktop Code tab, run on the user's local machine and have access to local files, MCP servers, skills, git, bash, worktrees, and the terminal, making them ideal for developers (e.g., code reviews, dependency audits, PR triage). These tasks are session-scoped and will stop running if the terminal closes, the computer sleeps, or after a 3-day expiry. The Desktop tasks utilize a self-improving loop pattern involving reading and writing a file to maintain memory between runs, allowing them to learn from past executions (e.g., fixing errors over time). Conversely, Cowork scheduled tasks run persistently in a VM sandbox in the cloud, utilizing connectors, plugins, and web search, making them better suited for non-developer tasks like summarizing emails, managing calendars, or running daily reports. The video also presents a matrix mapping task types (Business/Life vs. Code/DevOps) against time horizons (Short-lived vs. Long-running) to guide users on where to place specific tasks. Finally, a critical section addresses permission realities across all scheduling surfaces, noting that desktop tasks stall on permission prompts, CLI requires , and fixes often require manually granting 'always allow' permissions to tools first.