Under strict token limits, Large Language Models (LLMs) suffer from systematic retrieval biases when aggregating multi-document inputs. Batch concatenation triggers the "Lost in the Middle" effect, while sequential updates cause "Information Drift" where early context is discarded. This paper evaluates "Pyramid Aggregation"---a hierarchical tree reduction algorithm---against batch and sequential baselines using a representative lightweight long-context LLM. We benchmarked these approaches by extracting 32 distributed system errors under strict length and enumeration constraints. While the batch and sequential methods randomly pruned out the majority of micro-facts, the Pyramid method successfully consolidated all 32 error events into a balanced macro-summary by abstracting details into global failure categories while preserving key representatives. Finally, we discuss how the Antigravity Python SDK coordinates the asynchronous multi-agent orchestration.