Ralph Technique

Overview

The Ralph Technique is a loop that re-triggers an agent after each turn with the remaining task set. It exists because language models are lazy completers — they tend to address part of a task and then stop, especially when the task set is long or the context is complex. Ralph pushes them back in with accumulated context until the full task set is addressed or the agent explicitly signals it is stuck.

The name comes from the pattern of “ralphing” the model back to work — it’s a gentle but persistent push to complete what was started.

How It Works

def ralph(agent, task, context, max_iterations=10):
    """Push an agent until the task set is fully addressed."""
    for i in range(max_iterations):
        result = agent.run(task, context)

        if result.status == "done":
            return verify(result.output)

        if result.status == "stuck":
            return escalate_to_human(result.question)

        # Agent stopped but tasks remain -- push it back in
        remaining = result.incomplete_tasks
        context = context + result.partial_output
        task = f"Continue with the remaining tasks:\n{remaining}"

    return escalate_to_human("Max iterations reached")

The key mechanism: after each agent turn, check whether the full task set was addressed. If not, construct a new prompt with the remaining tasks and accumulated context, then re-trigger the agent. This loop continues until the agent either completes everything or explicitly signals it is stuck.

When to Use

Exploration mode. Parallelize multiple Ralph loops, each exploring a different approach to the same problem. The output is learnings, not code. Tech debt doesn’t matter because you will throw most of the code away and keep only the insights.

Completion mode. Well-defined task lists where you want to minimize human interruptions. The agent works through a checklist of changes, and Ralph ensures none are skipped.

When Not to Use

• When the task requires human judgment at each step — use Human-in-the-Loop instead.
• When the agent is heading in the wrong direction. Ralph amplifies misalignment: if the approach is wrong, pushing harder only makes it worse. Combine with a Hawk Agent to detect misalignment early.

Variants

1. Scratchpad + Task System (recommended). The agent maintains a structured task list as a scratchpad. On each turn, it updates the list with completed items and notes on remaining work. Ralph reads the scratchpad to determine what to push next.

2. Iteration loop with verification agent. A second agent reviews the output after each Ralph cycle before deciding whether to continue. This catches quality issues early but doubles the cost per iteration.

Trade-offs

Related Patterns

• Orchestrator-Worker — The orchestrator can use Ralph internally to ensure each worker completes its subtask.
• Human-in-the-Loop — Combine with Ralph so the agent can break out of the loop when genuinely stuck.
• Iterative Refinement — Similar loop structure, but focused on quality improvement rather than task completion.