The same work on a third of the tokens: tuning a fleet of coding agents | Alexander Marquardt

An engineer running a fleet of autonomous coding agents reduced token consumption by 3.3 times per changed line of code and increased merged pull requests by 4.4 times. The efficiency gains were achieved by measuring token usage from session transcripts, limiting each agent to a single task with a context ceiling enforced by a pre-tool-use hook, preventing prompt cache expiry during long waits, turning rules into deterministic code, and offloading routine fleet management to a program. Open article →