Agent loop

One turn of conversation runs an observe, act, inspect, adjust loop: the model sees tool results and decides the next step until the work is done or a bound stops it.

One turn, step by step

1. You send a message, optionally with queued follow-ups behind it.
2. The model may call tools. Reads run immediately; writes and shell pause for your approval in normal mode.
3. Each tool result returns as text and feeds the next decision. Failures arrive as Error: ... strings so the model can react rather than crash.
4. Transient failures retry. A denied call is never retried: the model replans instead.
5. Tokens, tool activity, and phase status stream live while this happens.

Bounds that end every turn

• 30 tool rounds by default (ATOM_MAX_TOOL_STEPS, clamped 5 to 100). Termination is structural.
• Compaction summarizes older turns near 83 percent of the verified window, so long work continues instead of dying.
• /clear drops history while keeping cumulative spend, for a deliberate fresh start.

Steering a running turn

• /queue <text> lines up a follow-up; /steer <text> injects into the running turn or sends when idle.
• Esc stops a running response. The footer shows esc stops while busy.
• Tab flips normal and yolo mid-session when the risk profile changes.

Read next

• Tools for the executors, caps, and background tasks behind each step.
• Permissions for the approval flow that gates writes.
• Compaction for how long runs survive the context limit.