Coding agents waste effort compressing rare logs, analysis finds

The paper's central claim rests on a specific technical assumption: that prefix caching already solves 98.6% of input token costs. This figure deserves scrutiny. The authors don't disclose whether their 600 trajectories come from systems already running prefix caching, or whether this percentage holds for agents operating without it (a common scenario in research labs and early-stage deployments).

This work directly challenges the optimization priorities embedded in recent coding-agent research. The Asana case study from last month showed that model routing and inference optimization can yield 76x cost reductions, but TurnCost suggests the field has been chasing compression and observation-filtering as proxies for efficiency when the real bottleneck is turn count itself. The earlier benchmarking papers (from October 7-8) flagged misalignment between evaluation and production workflows; TurnCost extends that critique into the optimization layer, arguing researchers are solving for problems that don't exist in real workloads.

If vendors deploying agents without prefix caching (or with older cache implementations) report that observation compression yields measurable cost savings within the next two quarters, the paper's generalizability is in question. Conversely, if the field shifts optimization focus toward turn reduction and away from token compression without pushback from practitioners, that signals the 98.6% figure has credibility.