The central argument is that tokens are not interchangeable merely because they cost the same. A conventional agentAn AI agent is a system that observes context, decides what to do, and takes actions through tools to pursue a goal. spends its whole budget executing a task, while a structured strategy assigns some tokens to other jobs that can improve the executor's decisions, evaluate its output, or retain lessons for later runs.
An advising strategy lets the executor request guidance while it works. A grading strategy checks each attempt against a defined rubricAgent evaluation tests whether an AI agent completes tasks correctly, consistently, and within its required boundaries. and repeats when the result falls short. A dreaming strategy reviews completed work, writes lessons to memoryAI agent memory is stored information that an agent can retrieve and use across steps, sessions, or changing contexts., and feeds those lessons into the next run. The examples span sales, customer support, and recruiting workflows.
A financial-analysis benchmark compares these strategies first with self-selected budgets and then with a fixed token ceiling. The fixed-budget results show a performance difference between plain execution and role-based strategies, while a stricter analysis counts only perfectly accurate outputs as useful and estimates the repeated-run token cost of reaching one.
The practical choice depends on the objective. Advising may be more token-efficientAI token efficiency measures how effectively a model or workflow turns consumed input and output tokens into useful results., while grading or dreaming may be preferable when the highest priority is increasing the share of fully correct runs. Larger systems can compose these roles into coordinated loopsAgent orchestration coordinates AI agents, tools, people, tasks, state, and control flow so a larger workflow reaches a verified outcome., but the benchmark does not imply that one strategy will dominate every domain.
Watch on YouTube
