Leaked 2021 Anthropic roadmap foresaw Claude Code, tool use, test-time compute
A July 2021 internal Anthropic research roadmap has surfaced after being filed in a US federal court in January as part of a copyright lawsuit and later rediscovered by the research community. Written by co-founder Jared Kaplan when Anthropic was about six months old with only a dozen or so staff, the document proposed a general AI assistant that could create, edit and explain software through natural language and eventually match "a software engineer colleague," naming tool use, reinforcement learning for tool use, model self-judging, test-time compute, trillion-parameter models and very large contexts, plus training data from code diffs, pull requests and debugging traces. It also preserves a July 16, 2021 internal debate: researcher Andy Jones asked whether to build a genuinely useful coding assistant or first prove tool use, GPT-3 first author Tom Brown suggested starting from code repo browsing, and CEO Dario Amodei decided to validate tool use in the short term and evolve into a strong coding assistant later; on safety, co-founder Sam McCandlish asked how to safely deploy tools without sandboxing.
The document indicates that the coding agents, tool use and self-judging ideas now central to the industry were already on Anthropic's research plan five years ago, and helps explain its long bet on code and safety.