Ralph Mode: We Put an AI Agent in a Loop and Let It Loose

Ralph Mode in action: company research, people search, audit, and iterative improvement.

What Is Ralph?

Ralph started as a bash loop:

Run an AI agent in a loop. Let it keep going until the task is done. The idea comes from Geoffrey Huntley, who spent months running Claude in a loop to build an entire programming language called CURSED. The technique is named after Ralph Wiggum from The Simpsons.

Each iteration starts with fresh context. The agent re-anchors from sources of truth on every pass, using file I/O as state rather than accumulating transcript. This prevents the context drift that kills long-running AI sessions.

We built this into Text2Resume. When you enable Ralph Mode, Rezzy stops waiting for you to approve each change and runs autonomously until the task is done.

How It Started

In early January, Vercel Labs shipped ralph-loop-agent, an experimental package built on AI SDK 6. It implements the loop concept with streaming, iteration tracking, and stop conditions.

We cloned the repo, added a streamAll() async generator to stream every iteration, and wired it up to our existing Rezzy tools. Text2Resume became one of the first production apps running Ralph mode.

Ralph Does Everything

Ralph has access to every tool in Text2Resume—the same 40+ tools as regular Rezzy mode. Content editing, layout adjustments, research, skills, projects, job imports, resume switching. Give him a task and he'll work through it until it's done.

For content optimization, Ralph often follows a research-first pattern: audit the resume, research the target company, improve weak bullets using what he learned, re-audit to verify improvement. But he adapts to the task. Ask him to redesign your layout and he'll iterate on typography, spacing, and colors until it looks right.

The forced context refresh on every iteration produces an interesting behavior: Ralph goes back and checks his own work. Because each iteration re-reads the resume from scratch, he notices if a previous edit didn't quite land. He'll revisit and refine.

This is especially visible with layout changes. Our layout tools give granular control over typography (20+ fonts, per-element sizing), colors (name, titles, accents, dividers), spacing (between sections, bullets, items), margins, column modes, and background fills. Ralph becomes a perfectionist here. He'll adjust heading weight, tweak section spacing, then come back two iterations later and nudge the margins because the balance is off.

This tinkering would be annoying in a single-pass agent. In Ralph mode, it's exactly what you want: an agent that keeps refining until the result looks right.

Anthropic's Plugin vs. Original Ralph

Anthropic released their own Ralph-inspired Claude Code plugin, but their approach differs. Their version uses a stop hook that blocks exit and re-feeds the prompt in the same session. The transcript grows. Context fills up. Failed attempts pollute future iterations.

This defeats the purpose. The original vision was fresh context per iteration, with file I/O as state instead of transcript. The agent should re-anchor from sources of truth on every pass. Anthropic's own long-context guidance says exactly this: agents must re-anchor from sources of truth to prevent drift.

Our implementation follows the original vision:

• Context management with summarization to prevent bloat
• Re-anchoring on every iteration from actual resume data
• Mutable working copy so the audit tool sees changes made during the session
• Verification callback that checks whether the agent accomplished something
• Stop conditions based on iteration count, cost, and completion signals

The Technical Details

Ralph Mode uses the same 40+ tools as regular Rezzy mode. Content tools (bullets, skills, experience, education, certifications, projects), layout tools (typography, colors, spacing, margins, column modes), research tools (Exa AI), job tools (import, search, select), resume management tools (import, list, switch). Plus one Ralph-specific tool:

• mark_complete signals that the agent believes the task is done

The RalphLoopAgent from AI SDK wraps the standard streamText call with iteration logic. It tracks cost, manages stop conditions, and runs a verification callback after each iteration.

The verification callback runs after each iteration. It checks whether the agent called mark_complete, whether it used tools, and whether the audit score improved. This prevents endless spinning and premature completion claims.

The UI

Ralph Mode shows progress differently than regular chat. Instead of individual tool cards waiting for approval, you see a task manager that tracks iterations.

Each iteration displays its status (in progress, completed, verified), tool count, and duration. Expanding shows which specific tools were called.

You can stop Ralph at any time. The current iteration finishes, then Ralph pauses. Like picking up an RC car mid-run, pointing it in a new direction, and letting it go. "Send & Resume" lets you redirect without losing context.

Constraints

Ralph Mode has guardrails built into the system prompt:

• Never remove entire experience entries (career history is protected)
• Max 3-5 skill changes per iteration
• No major restructuring without explicit request
• Target weak bullets only; leave A/B grades alone
• Preserve the user's voice

Ralph is limited to OpenAI models (GPT-5.2, GPT-5.1, GPT-5 Mini) for now. The iteration loop requires consistent behavior across many passes, and we've tuned it for OpenAI's tool-calling patterns.

Ralph Mode costs more credits than regular chat: 2x the base model cost upfront, plus per-iteration charges. The tradeoff is hands-off automation. You can start a Ralph run and return to a better resume.

A Typical Run

Here's a Ralph run on a software engineer resume targeting a specific job:

Total: about 2 minutes, four iterations, twelve tool calls. The resume is tailored to the target company with measurable quality improvement.

Try It

Ralph Mode is live in Text2Resume. Select a target job, enable Ralph Mode, and let it run while you do something else.