Coca-Cola Compresses Tool Outputs Into Latents

Coca-Cola makes a compelling case that agents need decision-relevant context, not human-readable transcripts—but its generality beyond web search remains unproven.

• –Action-preserving training aligns compressed and full-context next-action distributions, a more useful objective than reconstructing tool text.
• –The 128-token budget is configurable; experiments show performance changes as the latent-token count increases.
• –Compressing tool observations attacks a major agent bottleneck, since repeated tool outputs can dominate long-horizon contexts.
• –The open-source repository includes training, evaluation, and comparisons against Vanilla, LLMSummary, and LongLLMLingua.
• –Model-specific embedding integration and training requirements may limit plug-and-play adoption across agent backbones.