Arnold

Arnold — Build Intelligent Pipelines

Arnold is a tool for building intelligence systems out of many coordinated models. Today you can experience one slice of that through its first tool, **Megaplan** — a planning and execution harness for software. More on building and sharing your own pipelines soon.

---

Megaplan - a pipeline for cost-efficient advanced planning and execution


It breaks building software into structured, independently-checked phases — making intelligent-but-unreliable LLMs systematically robust, and letting each phase run on the cheapest model that can do it well.

Megaplan's philosophy

Two ideas, and they reinforce each other.

**1. Structure makes LLMs robust.** Modern LLMs are highly intelligent but systematically unreliable. Left to their own devices they skip steps, miss concerns, rubber-stamp their own work, and research sloppily. Megaplan breaks the whole process into explicit stages — `prep`, `plan`, `critique`, `gate`, `revise`, `finalize`, `execute`, `review` — each scoped and equipped so the model does one thing well, and each checked by a *separate* pass instead of the model grading itself. That structure is what turns raw intelligence into something that can deliver a real-world sprint end-to-end, reliably.

**2. Use the cheapest capable model per component.** Premium closed models (Claude, GPT/Codex) are overkill for the vast majority of software tasks; open models that are **~40× cheaper** — primarily **DeepSeek v4-pro** — do most of the work just as reliably. The goal is to use the cheapest model that can *reliably* perform a given task, and to spend premium budget only where it actually changes the outcome.

The two connect: the same decomposition that makes the process robust is what lets you price it. Once work is broken into stages and tasks, each routes to the cheapest model that can handle it — premium models reserved for two jobs: **adjudicatin