Parallel test-time scaling treats each rollout as an independent candidate answer. For long-horizon search agents, however, these rollouts also produce rich traces of tool calls, retrieved evidence, and partial reasoning states. Intermediate failures such as reasoning drift are difficult to self-correct within a single rollout, yet useful corrective signals often already exist in other live rollouts before termination. This creates an opportunity to turn cross-trajectory information into online correction during execution, rather than harvesting it only after rollouts terminate. However, simple broadcast is unsafe because partial states can be redundant, irrelevant, or risky for the receiving trajectory. We propose OPTA, a live aggregation framework based on checkpoint-level synchronization: each rollout maintains a private report, while a controller distills a typed global shared state, and injects only relevant and useful feedback into each trajectory, preserving its exploration identity. On BrowseComp-Plus and DeepSearchQA across four LLM backbones, OPTA is consistently the strongest test-time scaling method, lifting Qwen3-30B on BrowseComp-Plus from 37.9% to 59.6%. Our results show that OPTA is a practical route toward timely and selective test-time scaling for long-horizon LLM search agents.