Tokens Should Have Jobs — Katelyn Lesse & Angela Jiang, Anthropic

Tokens in agentic systems are not fungible; instead, assigning them distinct functional roles—advising, grading, and dreaming—significantly improves task outcomes. While increasing total token budget generally enhances performance, specific strategies yield superior "alpha" for complex tasks like financial analysis. Advising allows executors to seek guidance during execution, grading ensures output meets strict rubrics, and dreaming enables agents to store learnings in memory for future iterations. Experiments demonstrate that while basic execution requires massive token expenditure to achieve perfect accuracy, specialized strategies optimize for either token efficiency or high-reliability results. Businesses should select strategies based on whether they prioritize cost-effectiveness or the necessity of error-free outputs. Integrating these primitives into agentic architectures allows for the construction of sophisticated, self-improving systems capable of handling dynamic, real-world challenges with greater precision and efficiency.