Autonomous Loop Patterns

Concept Card

What it is: Autonomous loop patterns are reusable control shapes for running an agent repeatedly until an explicit stop condition is reached. They define how a session continues, stops, checkpoints, respawns, and avoids runaway behavior.

Mental model: A loop has five primitives: a trigger, a worker session, a progress signal, a stop condition, and a safety cap. Ralph Wiggum puts the stop condition in a Stop hook and completion word. Manage-loop puts it in a continuation signal. Grind-loop puts it in an external supervisor that respawns fresh sessions based on status files and exit outcomes.

Why it exists: Without named loop patterns, agents invent ad hoc respawn logic that has no checkpoint, no cap, and no clear owner for recovery. These patterns let the system choose the smallest loop that can safely finish the work.

What it is NOT: It is not command selection, task execution, inter-agent messaging, or budget steering. Those are handled by ai-coding-agents, task-execution, agent-to-agent, and agent-control.

Adjacent concepts: Hook patterns provide lifecycle intercepts, loop infrastructure provides checkpoint mechanics, agent control provides steering and abort signals, and observability provides evidence that a loop is progressing.

One-line analogy: Autonomous loop patterns are autopilot modes: each mode keeps moving, but each needs instruments, altitude limits, and a clear handoff back to the pilot.

Common misconception: A loop is not safer because it can run forever. A safe autonomous loop is defined by when it stops, what it persists, and how it proves progress between iterations.

Domain Context