Leaked 2021 Anthropic Roadmap Sketched Coding Agents, Tool Use and Long Context

The document, titled "Concise Research Roadmap," runs only three or four pages and was written by Anthropic co-founder Jared Kaplan. It carries annotations from the company's researchers. Co-founder Tom Brown called repository browsing a simple, high-value entry point, while Nelson Elhage, a researcher in systems and interpretability, predicted that automatically locating type definitions inside a project would be a highly valuable feature for users.

Facing a dispute over whether to pursue the strongest coding assistant or a tool-use demonstration, chief executive Dario Amodei set the order: demonstrate tool use in the short term, build the strongest coding assistant in the long term.

Alignment scientist Amanda Askell raised a deeper concern in the document. If alignment work were confined to natural language while capability research concentrated on code, she wrote, the two would drift apart. Safety and capability, she argued, have to be tied together at every stage.

At the time, Anthropic was six months old and had raised $124 million in total. ChatGPT was 16 months away and the first Claude model 20 months away, and much of the industry still questioned whether large models were a transitional product. The group that left OpenAI was already planning work on extremely high GPU utilization, long context, reinforcement learning for tool use and a framework for assessing social impact.

The technical core of the roadmap was a large generative model with long-term memory, fine-tuned through reinforcement learning that combined automated rewards with human feedback. The first carrier for the approach was to be a conversational software engineer assistant able to create, edit and explain software, write functions, find bugs, design APIs and discuss code style and engineering practice, reaching the level of a human engineering colleague — the prototype of what is now Claude Code. Scaling came first in the plan. Kaplan, first author of the Scaling Laws paper, brought that view from OpenAI and was already considering models of one trillion parameters and beyond, along with GPU utilization, early adaptation to TPU and Trainium hardware, packaging block sparse attention so non-specialists could configure hyperparameters, and protecting model weights as a core asset.

The plan also covered pretraining data. Anthropic intended to train the best-performing language-and-code base model in the industry as a foundation for later fine-tuning, and to build full-pipeline datasets covering not only natural language and code but also diffs, pull requests, debugging logs and evaluation results from code execution. Data filtering, migration, fine-tuning and diversity optimization were all to be handled in-house.

On alignment, the document called for training models that not only answer questions but also offer help, adjust behavior based on feedback and ask for clarification when user intent is ambiguous. It listed scenarios where alignment fails, classified them by difficulty and studied how alignment behaves as models grow. Reward models were to be cross-validated on compliant human-feedback data, static datasets such as Stack Exchange and adversarially generated data, with the results distilled into a general methodology quantifying how much data and compute different alignment goals require. The article links this work to Claude's early adoption in finance, healthcare and legal settings.

Interpretability was treated as a long-term investment. Christopher Olah, a leading researcher in the field, was to break down neurons, attention heads and internal functional circuits, and to connect data and tool research with interpretability work so that arithmetic reasoning, multi-turn dialogue and code generation could be understood at the level of mechanism. Reinforcement learning for tool use was listed as a separate research direction, with information retrieval and printf debugging as early tasks, to test whether tool interaction beats supervised learning in scaling behavior and compute efficiency. Amodei also proposed a natural language editor that could revise documents in a Google Docs-like environment, arguing that the tools and infrastructure behind the two kinds of editors overlap. The underlying question was whether a model can move from generating answers to completing complex tasks with tools.

Safety ran through the plan. Sam McCandlish asked how non-sandboxed tools could be deployed safely, and Amodei warned that tools able to execute code at deployment carried very high risk. The sandbox isolation and permission interception now debated in the agent field appear in those early discussions. Anthropic also wanted to position itself as a model for AI ethics research, working with academia and government on independent evaluations.

What followed tracked the document. Half a month after ChatGPT was released at the end of 2022, Anthropic published its Constitutional AI approach, in which a model critiques and revises another model's answers against a set of behavioral principles, addressing the cost, scaling limits and annotator bias of reinforcement learning from human feedback. In March 2023 the first Claude was released to partners, seen as noticeably more polite than GPT-3.5. In May the context window went from 9,000 to 100,000 tokens, about 75,000 English words, while GPT-4 then supported up to 32,000. September brought the first Responsible Scaling Policy, which graded risks and tightened controls as capability rose, and in October a Google-led round valued Anthropic at $18.4 billion. In March 2024 the Claude 3 family added multimodal capability, and the flagship Opus supported 100,000 tokens by default and outperformed GPT-4 on long-text evaluation. Claude 3.5 Sonnet matched GPT-4o in June, and an October upgrade surpassed all closed-source competitors on the SWE-bench Verified coding benchmark.

Five years on, Claude can process one million tokens at once, the flagship line has reached the Opus 5 generation, and safety training has moved from human-labeled feedback to Constitutional AI and then to hybrid methods. Readers responding to the leaked document said the agents the industry is now racing to build had been on Anthropic's to-do list, while others argued that such foresight is not unusual among frontier researchers and that executing the plan without drifting was the harder achievement.