AI agent token optimizations backfire, increasing costs by up to 82%

The Altair project team tested three token-reduction strategies on their open-source AI agent. Reducing system prompt length, aggressively clearing old context, and loading tools individually all increased costs per task by up to 82%. The optimizations failed because AI agent pricing depends more on request steps and cache usage than individual request size. Shorter prompts caused the agent to take 57% more steps per task, negating token savings. Tool description truncations also caused task failures that didn't occur with full descriptions.