Tokens Should Have Jobs: Optimizing Agentic Strategies

The Shift from Execution to Strategy

Most agentic systems treat token budgets as a fungible resource, assuming that simply increasing the budget leads to better outcomes. Katelyn Lesse and Angela Jiang argue that tokens should be assigned specific jobs to improve performance. By splitting a fixed token budget between an executor and other specialized roles, developers can achieve higher accuracy and reliability without necessarily increasing total spend.

Core Agentic Strategies

• Advising: An executor performs the primary task but can query a secondary adviser agent to validate steps or flag issues mid-process.
• Grading: A grader agent evaluates the executor's output against a predefined rubric. If the output fails to meet the criteria, the system forces the executor to iterate until it achieves a passing result.
• Dreaming: A dreamer agent inspects the executor's transcripts and findings, writing key learnings into a memory store that subsequent runs can access to improve performance over time.

Performance and True Cost

When evaluated on a financial analysis benchmark, a pure execution strategy achieved 76% accuracy on a 600,000-token budget. By reallocating that same budget to an advising strategy, accuracy increased to 89%.

When applying a strict requirement for 100% accuracy—where partial results are considered useless in a financial context—the baseline execution strategy passed only 42% of the time. This necessitated an average of three runs to achieve a single perfect result, resulting in a true cost of 1.8 million tokens. More complex strategies like grading and dreaming significantly reduced this total cost by increasing the probability of a successful run per attempt. Developers should choose strategies based on whether they are optimizing for raw token efficiency or absolute reliability.