Better AI agent systems scale by remembering useful feedback, not by spending more compute.
Rohan Paul Twitter · Rohan Paul (@rohanpaul_ai) · 2026-06-01
A research paper argues that AI agent systems improve most effectively by storing and reusing useful feedback rather than spending more compute, challenging the assumption that token counts or API call volume signal meaningful progress.
Appears in
Extraction
Topics: ai-agentsscalingfeedback-mechanismscompute-efficiency
Claims
- AI agent systems scale more effectively by remembering and reusing useful feedback than by increasing compute expenditure.
- Treating token counts, API call counts, or dollar costs as proxies for agent progress is a common and misleading mistake.
- Two agent runs can spend identical compute budgets while achieving very different outcomes, making resource consumption an unreliable performance metric.
- The quality of feedback stored and retrieved, not volume of compute consumed, is the meaningful signal for agent improvement.
Key quotes
Better AI agent systems scale by remembering useful feedback, not by spending more compute.
The simple mistake is to count tokens, calls, or dollars as if they were all evidence.
2 runs can spend the same budget while [achieving very different outcomes].