Input tokens consume 54% of usage in agentic coding, revealing costly 'communication tax'

rseroter · x · 2026-08-19

Research reveals that input tokens account for 54% of total usage in LLM-based multi-agent software engineering systems. A costly "communication tax" emerges as AI agents repeatedly pass full contexts back and forth for dialogue, code review, and iteration. This constant context passing drives up financial costs, energy use, and environmental impact. The author notes that while most research focuses on capabilities, understanding token consumption patterns is critical as these systems scale from demos to real workloads.

Original post →

More from coding & agent

coding & agent channel →