The authors profiled token use of Claude Code, Codex, and GitHub Copilot CLI on 60 SWE-Bench tasks across 6 repositories, testing three fixed strategies from popular plugins: command-output compression (Rust Token Killer), semantic navigation (Serena), and structural context injection (Aider's repository map).
Environment-aware controller cuts coding agent token costs without hurting success
TokenEconomizer learns per-repository policies from completed traces, pruning unused tools and interrupting failure loops to save 6.4 to 30.9 percent of tokens across three commercial coding agents.
Independent
Aaron J. Li · Weijia Xu · Chandan Singh · Zixiao Chen · Jeevana Priya Inala · Charlotte Siska · +5 more
Research Digest··2 min read
The authors evaluate coding agents over sequences of repository tasks and find that fixed cost-saving strategies work unreliably, so they build TokenEconomizer, an external controller that learns from past executions.
§