Anthropic's 2021 internal research roadmap surfaces in copyright suit, sketching Claude Code years early

A document marked "top secret" inside Anthropic has surfaced through a copyright lawsuit, and it reads like a map of the products the company ships today. Chinese tech outlet QbitAI reported on Oct. 15 that the file is an internal research roadmap written in July 2021 — only two or three pages long — and that it anticipated Claude Code, tool use, and model self-judging years before those ideas became industry shorthand.

The timing is the striking part. Anthropic had spun out of OpenAI only months earlier and employed roughly a dozen people, GPT-3 was still the reference point for most observers, and the common imagination for AI and programming stopped at code autocomplete.

The roadmap opens with an unusually concrete definition of its long-term goal — safe, beneficial AGI — anchored to a prototype: a general AI assistant that talks with people in natural language and creates, edits, and explains software. The document lists answering general questions, writing functions, finding bugs, designing APIs, and discussing code style as expected skills, and states that the assistant should demonstrate the full capabilities of a software engineering colleague.

On implementation, the file preserves an internal disagreement. Researcher Andy Jones asked whether the team should first build a genuinely useful coding assistant or first prove that tool calling worked at all; Tom Brown, first author of the GPT-3 paper, suggested starting from repository browsing. CEO Dario Amodei settled it: validate tool use in the short term, evolve into a strong coding assistant in the long term.

The pretraining and alignment sections reach just as far. The team wrote down plans to train models of 1 trillion parameters and above, and to explore very large contexts or other mechanisms that could hold vast amounts of memory and experience. The alignment chapter includes model self-judging — using a model's own judgement as a feedback signal — and mentions test-time compute, though the report notes these sat in the context of alignment and discriminator research rather than today's o1-style inference scaling.

Safety takes up a large share of the document. Co-founder Sam McCandlish asked how the company could safely deploy tools without sandbox isolation, and Amodei replied that he meant tools that execute code at deployment time. Those annotations map closely onto today's debates about agent privilege escalation, environment escape, data poisoning, and prompt injection.

The file also records internal caution. Amanda, described in the report as Anthropic's in-house philosopher, warned that doing alignment research on natural language while advancing capability mainly on code would limit the validity of alignment tests and could turn alignment work into an island. She proposed a company principle: capability and alignment research should advance in proportion, unless there is a clear reason to separate them.

QbitAI ends with its own reading — Anthropic spent five years on high-quality code signals, reinforcement learning, and enterprise workflows, while OpenAI spread across image generation, video, voice, and search. For readers, the more useful question is how much of a 2026 product line can be traced to a 2021 planning document, and whether that changes how the industry thinks about sequencing strategy and capability.