A separate study pruned 41–46% of reasoning tokens with little loss in task performance. Clean up the context before you add skills or subagents.
From the Field Notes:
Trimming what goes into an agent's context beats adding more harness machinery
In one minute
Surge AI found that raising Opus 5.5's thinking effort gained more than any harness it tested, including one with 37 skills and 17 agents. EnvOpt optimized the agent's information environment, cut Claude Haiku 4.5 cost by 72.7% and raised its pass rate.