Large language model (LLM) agents usually append raw tool outputs to their context during multi‑step interactions. As the number of steps grows, long and noisy observations cause substantial context overhead and dilute the agent's focus on task‑relevant information. Existing token‑pruning and summarization techniques are limited by discrete text representations, restricting achievable compression ratios. We introduce Coca‑Cola, a continuous latent context compression model that challenges the assumption that agents must consume tool observations as explicit text tokens. Coca‑Cola compresses tool outputs into latent tokens and embeds them directly into the agent's input embedding space. To ensure that compression preserves decision‑making ability, we propose Action‑Preserving Training, which aligns the agent's next‑action distributions under full and compressed observations, guiding the compressor to retain only action‑sufficient information. Experiments with web‑search agents on four QA benchmarks across two model families show that Coca‑Cola reduces token consumption by 70.01%‑89.24%, cuts end‑to‑end runtime by 24.45%‑55.97%, and consistently improves task performance. The code is released at https://github.com/THU-Agent/Coca-cola.
Review: Coca‑Cola demonstrates that latent‑space compression can dramatically boost the efficiency and effectiveness of LLM agents, opening a promising direction for future tool‑augmented and multimodal systems.