The Fallacy of Fungible Tokens
Most teams attempt to improve agent performance by simply increasing the token budget. This assumes that all tokens are fungible and that "more compute" is the only lever for better outcomes. However, research from Anthropic demonstrates that for a fixed budget, the allocation of tokens to specific roles (jobs) significantly outperforms a brute-force execution strategy.
Three Functional Agent Strategies
Rather than having every token perform the same execution task, builders should implement specialized strategies using these primitives:
- Advising: Splits the system into an executor and an adviser. The executor calls out to the adviser mid-task to verify steps, which is particularly effective for complex workflows like sales or financial analysis.
- Grading: Defines a rubric of what "success" looks like upfront. A grader agent evaluates the executor's output against this rubric; if it fails, the executor iterates until it meets the criteria.
- Dreaming: An agent inspects past transcripts and findings, writing these learnings into memory. This memory is then used by the executor in subsequent runs to improve performance over time.
Optimizing for Real-World Reliability
In domains like financial analysis, an 80% accurate result is often useless because errors (e.g., incorrect cost numbers) cannot be tolerated. When re-evaluating performance based on a "perfect run" requirement, the baseline execution strategy had a 42% pass rate, requiring an average of three runs (1.8 million tokens) to achieve one usable result.
By contrast, strategies like advising and grading reach higher reliability thresholds with significantly lower total token costs. The choice of strategy should depend on the business goal: prioritize token efficiency by using advising, or prioritize reliability by using grading or dreaming. These strategies are not mutually exclusive and can be combined into complex, multi-agent architectures to solve dynamic problems.