Agent Loop

The Three-Step Cycle

1. Reason

The attacker LLM produces a reasoning trace — chain-of-thought planning the next move.

2. Tool-Call

Select and invoke a tool: probe, mutate, escalate, summarize, finish.

3. Observe

The tool output is appended to conversation history. The agent sees the response, errors, and metadata.

Exit Conditions

• finish() — Attack complete
• ask_operator() — Pause for human guidance (TUI only)
• Loop cap — The agent loop runs up to 25 iterations per round (max_iters in agent/loop.py); one-shot mode caps total rounds with --rounds (default 12)
• User interrupt — Ctrl+C

Concurrency