Core Features
Agent Harbor provides several key capabilities that differentiate it from running AI coding agents directly.
Sandboxed Execution
Agents run in isolated environments where they cannot:
- Access sensitive files (SSH keys, credentials)
- Modify your host filesystem directly
- Conflict with other running agents
- Access the network without explicit permission
Every change is captured in a copy-on-write overlay and can be reviewed before applying.
Time Travel & Forking
Rewind to any point in an agent’s session and branch off:
- Filesystem snapshots at logical boundaries
- Visual replay of terminal output
- Create new branches with modified instructions
- Compare different execution paths
Works with native CoW filesystems (ZFS, Btrfs) or the cross-platform AgentFS.
Learn more about Time Travel →
Recording & Replay
Every session is captured for later analysis:
- Full PTY stream with exact timing
- Snapshot markers for navigation
- Brotli compression for efficient storage
- Extract branch points for intervention
Parallel Agent Execution
Launch multiple agents on the same task:
- Each agent gets its own filesystem branch
- No conflicts between concurrent sessions
- Compare results and cherry-pick the best solution
- Scale horizontally for complex tasks
Filesystem Snapshot Providers
| Provider | Platform | Best For |
|---|---|---|
| ZFS | Linux | Production servers with ZFS pools |
| Btrfs | Linux | Development machines with Btrfs |
| AgentFS | All | Cross-platform consistency |
| Git | All | Universal fallback |
Terminal Integration
Works with your existing terminal setup:
- tmux, zellij, screen
- kitty, iTerm2, WezTerm
- Vim, Neovim, Emacs terminal modes
No special terminal required — Agent Harbor integrates with what you already use.