GitHub Analysis Reveals 19-62% Token Reductions by Eliminating Unnecessary LLM Calls

The Structural Shift in Agentic Workflows

A May 7 analysis by GitHub of five production agentic workflows has revealed that the cheapest token is the one you never send. The engineering team found that token reductions of 19 to 62 percent came not from better prompting techniques, but from removing large language model calls entirely for steps that did not require reasoning. The key finding is structural rather than algorithmic, highlighting that most agent turns are deterministic data-gathering steps that do not need an LLM in the first place.

Optimizing Context and Eliminating Agent Turns

To achieve these efficiency gains, the engineering team focused on pruning unused Model Context Protocol tools, which saved 8 to 12 KB of schema context per call. Additionally, replacing GitHub Model Context Protocol calls with direct command line interface commands eliminated entire agent turns, dramatically reducing overhead and improving execution speed.