AI agents with autowt
autowt provides powerful capabilities for working with command-line AI agents like Claude Code, enabling you to monitor their activity and manage multiple development sessions effectively. By creating isolated environments for each task, autowt prevents agents from interfering with each other while providing real-time visibility into their status.
Why autowt for AI agents?
Running multiple AI agents in a single directory creates chaos. One agent might overwrite files from another, or you might lose track of which changes belong to which task. autowt solves these problems through context isolation, where each agent operates in its own worktree with its own directory. This eliminates context pollution and file conflicts entirely.
The parallel execution capabilities allow you to spin up multiple agents on different tasks simultaneously. One agent can write tests while another develops a new feature, all without interference. When you need to switch between tasks, autowt preserves the state of each worktree so you can pick up exactly where you left off. Once an agent completes its task, you can easily clean up its worktree to keep your project organized.
Setting up agent monitoring
autowt requires hooks to monitor agent activity. Install them using the interactive setup command:
You'll be prompted to install hooks at the user level (affecting all projects), the project level (current project only), or have the JSON configuration printed for manual installation. User-level installation is recommended for consistent monitoring across all your projects.
To check which hooks are currently installed and whether they need updating:
Understanding agent status
Once hooks are installed, autowt automatically tracks Claude Code agents running in your worktrees through integration with Claude's hook system. This provides real-time visibility into what each agent is doing.
When you run autowt ls, agent status indicators appear alongside terminal session markers:
Worktrees:
→ ~/dev/my-project (main worktree) main ←
~/dev/my-project-worktrees/feature-new-ui @C? feature-new-ui
~/dev/my-project-worktrees/hotfix-bug @C+ hotfix-bug
~/dev/my-project-worktrees/refactor-auth @ refactor-auth
The status indicators reveal agent activity at a glance. C? indicates a Claude agent waiting for your input, while C+ shows an agent actively working on a task. C- means the agent is idle after recently completing work, C! signals notifications or errors, and C* indicates a subagent task has completed. A simple @ shows an active terminal session without an agent.
Live monitoring dashboard
For comprehensive agent oversight, the live monitoring dashboard provides real-time updates:
This interactive terminal interface shows all your worktrees with automatic status refreshes every few seconds. The dashboard displays live status updates, activity timelines showing when each agent was last active, and provides quick switching to any worktree with a simple keypress.
Navigation within the dashboard uses intuitive keyboard shortcuts. Use the arrow keys or j/k to move between worktrees, press Enter to switch to the selected worktree, or press w to jump directly to the first agent waiting for input. Press r to manually refresh status and q to quit the dashboard.
Agent workflows
The most effective pattern for AI agent development involves parallel feature development using the --after-init flag to launch agents immediately after worktree creation.
Assign different types of work to appropriate agents. For feature development, create a worktree and launch Claude directly:
This creates the feature branch worktree and immediately starts the Claude agent in a new terminal. For documentation tasks where different AI models excel, you might use:
autowt docs/api-reference \
--after-init 'gemini "Write comprehensive documentation for the new API endpoints, including examples."'
When you need to work on sensitive code without agents, simply create worktrees without the after-init command:
This approach creates three parallel development streams, each in an isolated environment with dedicated terminal sessions for monitoring progress independently.
Automating environment setup
The --init flag or autowt configuration files can automate agent environment preparation. Setting an init script in your .autowt.toml file ensures consistent setup:
With this configuration, dependency installation runs automatically in every new worktree. Combining init scripts with after-init commands creates fully automated workflows where the init script prepares the environment before the agent begins its task.
Smart agent switching
autowt extends switching capabilities with agent-aware options for efficient session management. When multiple agents are running, quickly jump to those needing attention:
This command switches to the first agent waiting for input, presenting an interactive choice if multiple agents need attention. To jump to the most recently active agent:
These commands integrate seamlessly with the monitoring system to provide intelligent navigation based on actual agent activity rather than manual tracking.