Overview
Agents are Claude Code SDK instances running within a workspace’s git worktree. You see every message, tool call, file edit, and cost in real time.Starting Agents
Start an agent by writing a prompt in a workspace. You can run multiple agents per workspace for parallel tasks. Each message supports per-message overrides:| Option | Values | Description |
|---|---|---|
| Model | Default, Opus, Opus 1M, Sonnet, Haiku | Override the default model |
| Effort | Low, Medium, High, Max | Control thoroughness |
| Thinking | On/Off | Extended thinking mode |
| Fast Mode | On/Off | Faster output |
| Context 1M | On/Off | Use 1M context window |
| Plan Mode | On/Off | Step-by-step with approval |
Lifecycle Controls
| Action | Description |
|---|---|
| Pause | Suspend the agent, preserving session |
| Resume | Continue from where it paused |
| Kill | Terminate immediately |
| Resume All | Resume all paused agents across workspaces |
| Rerun All | Create fresh agents for all workspace slots |
Message Queue
When an agent is busy, messages are queued:- Queued messages are persisted to SQLite (crash-safe)
- Auto-delivered when the agent becomes available
- Can be cancelled or force-delivered
- Each queued message preserves its model/effort/thinking options
Plan Mode
When enabled, agents propose a step-by-step plan before executing:- Agent generates plan steps
- You review and approve (or edit)
- Agent executes approved steps
- Progress tracked per-step
Session Resumption
Agent sessions persist across app restarts:- Session IDs stored in state
- On restart, agents with valid sessions resume automatically
- Agents without sessions are marked as errored with “Interrupted by app restart”
Token Tracking
Per-agent tracking includes:- Input/output token counts
- Cache creation and read tokens
- Cost in USD
- Context window fill percentage
- Daily aggregate usage