agentic-loop-autonomy-ladderlisted

# The Agentic Loop Autonomy Ladder **Source**: AI Builder Club, "The 4 Types of Agentic Loops" and "Loop Engineering Guide (2026)" — describes Claude Code's own `Auto` mode, `/goal`, `/loop`, and Routines/scheduled-task features as concrete implementations of each rung. ## The four types, by what they hand off Each type hands off strictly more of the loop than the one before it — the source: *"Each hands off strictly more of the loop than the one before it, which is why they form a ladder rather than a menu."* Not four alternatives to pick between freely; a progression. 1. **Turn-based — hand off tool approval.** The agent approves its own tool calls within one turn without interruption (e.g., Claude Code's Auto mode), but still decides when to *stop* based on its own judgment — the weakest verifier available. Fine for supervised, interactive work; vulnerable to doom-looping once left unattended. 2. **Goal-based — hand off the stop condition.** A measurable completion condition is set explicitly (e.g., Claude Code's `/goal`); after each turn, an evaluator (ideally a small, fast, independent model, per the source) checks whether the condition holds. This is the pivotal rung — the first one where genuinely unsupervised operation becomes defensible, because a real verifier now governs stopping instead of the agent's own judgment. 3. **Time-based — hand off the trigger to a clock.** The loop reruns on an interval (e.g., `/loop`), reacting to elapsed time rather than the agent